© CFA Institute. For candidate use only. Not for distribution.

READING

7
Machine Learning
by Kathleen DeRose, CFA, and Christophe Le Lannou
Kathleen DeRose, CFA, is at New York University, Stern School of Business (USA).
Christophe Le Lannou is at dataLearning (United Kingdom).

LEARNING OUTCOMES
Mastery The candidate should be able to:

a. distinguish between supervised machine learning, unsupervised

machine learning, and deep learning;
b. describe overfitting and identify methods of addressing it;
c. describe supervised machine learning algorithms—including
penalized regression, support vector machine, k-­nearest neighbor,
classification and regression tree, ensemble learning, and random
forest—and determine the problems for which they are best
suited;
d. describe unsupervised machine learning algorithms—including
principal components analysis, k-­means clustering, and
hierarchical clustering—and determine the problems for which
they are best suited;
e. describe neural networks, deep learning nets, and reinforcement
learning.

INTRODUCTION
Investment firms are increasingly using technology at every step of the investment
1
management value chain—from improving their understanding of clients, to uncov-
ering new sources of alpha, to executing trades more efficiently. Machine learning
techniques, a central part of that technology, are the subject of this reading. These
techniques first appeared in finance in the 1990s and have since flourished with the
explosion of data and cheap computing power.
This reading provides a high-­level view of machine learning (ML). It covers a
selection of key ML algorithms and their investment applications. Investment prac-
titioners should be equipped with a basic understanding of the types of investment
problems that machine learning can address, an idea of how the algorithms work,

© 2019 CFA Institute. All rights reserved.

 © CFA Institute. For candidate use only. Not for distribution.
464 Reading 7 ■ Machine Learning

and the vocabulary to interact with machine learning and data science experts. While
investment practitioners need not master the details and mathematics of machine
learning, as domain experts in investments they can play an important role by being
able to source appropriate model inputs, interpret model outputs, and translate out-
puts into appropriate investment actions.
Section 2 gives an overview of machine learning in investment management.
Section 3 defines machine learning and the types of problems that can be addressed
by supervised and unsupervised learning. Section 4 describes evaluating machine
learning algorithm performance. Key supervised machine learning algorithms are
covered in Section 5, while Section 6 describes key unsupervised machine learning
algorithms. Neural networks, deep learning nets, and reinforcement learning are
covered in Section 7. The reading concludes with a summary.

2 MACHINE LEARNING AND INVESTMENT

MANAGEMENT
The growing volume and exploding diversity of data, as well as the increasing eco-
nomic value of insights extracted from these data, have inspired a resurgence in data
science. This newly evolving field combines mathematics, computer science, and
business analytics. It also strikes out in a new direction that relies on learning—from
basic learning functions that map relationships between variables, to advanced neural
networks that mimic physical processes that absorb, order, and adapt to information.
Machine learning has theoretical and practical implications for investment man-
agement. For example, machine learning could potentially reshape received wisdom
about asset risk premiums and reconfigure investment management business processes.
Large data sets and learning models are already affecting investment management
practices—from client profiling, to asset allocation, to stock selection, to portfolio
construction and risk management, to trading.
Machine learning applications are at each step of the asset and wealth management
value chain. Chatbots answer basic retirement savings questions, learning from their
interactions with investors. Machine learning methods can be used to generate alpha
signals used in security selection by creating a non-­linear forecast for a single time
series, by deriving a forecast from a suite of predefined factors, or even by choosing
input signals from existing or newly found data. For example, researchers using textual
analysis have found that year-­over-­year changes in annual (10K) and quarterly (10Q)
filings, particularly negative changes in the management discussion and risk sections,
can strongly predict equity returns.
Machine learning methods can help calculate target portfolio weights that incor-
porate client restrictions and then dynamically weight them to maximize a Sharpe
ratio or better estimate the variance–covariance matrix by using such techniques
as principal components analysis, which reduces the number of variables needed to
explain the variation in the data. Research suggests that machine learning solutions
outperform mean–variance optimization in portfolio construction. Machine learning
techniques are already creating better order flow management tools with non-­linear
trading algorithms that reduce the costs of implementing portfolio decisions. These
developments have caused an evolution in automated tools, automated processes,
and algorithmically automated businesses, like robo-­advising.
 © CFA Institute. For candidate use only. Not for distribution.
What Is Machine Learning? 465

WHAT IS MACHINE LEARNING?

We now discuss some fundamental concepts of machine learning, including a defini-
3
tion of machine learning and an overview of key types of machine learning, such as
supervised and unsupervised ML.

3.1  Defining Machine Learning

Statistical approaches and machine learning techniques both analyze observations
to reveal some underlying process; however, they diverge in their assumptions, ter-
minology, and techniques. Statistical approaches rely on foundational assumptions
and explicit models of structure, such as observed samples that are assumed to be
drawn from a specified underlying probability distribution. These a priori restrictive
assumptions can fail in reality.
In contrast, machine learning seeks to extract knowledge from large amounts of
data with no such restrictions. The goal of machine learning algorithms is to automate
decision-­making processes by generalizing (i.e., “learning”) from known examples to
determine an underlying structure in the data. The emphasis is on the ability of the
algorithm to generate structure or predictions from data without any human help. An
elementary way to think of ML algorithms is to “find the pattern, apply the pattern.”
Machine learning techniques are better able than statistical approaches to handle
problems with many variables (high dimensionality) or with a high degree of non-­
linearity. ML algorithms are particularly good at detecting change, even in highly
non-­linear systems, because they can detect the preconditions of a model’s break or
anticipate the probability of a regime switch.
Machine learning is broadly divided into three distinct classes of techniques:
supervised learning, unsupervised learning, and deep learning.

3.2  Supervised Learning

Supervised learning involves ML algorithms that infer patterns between a set of
inputs (the X’s) and the desired output (Y). The inferred pattern is then used to map
a given input set into a predicted output. Supervised learning requires a labeled data
set, one that contains matched sets of observed inputs and the associated output.
Applying the ML algorithm to this data set to infer the pattern between the inputs
and output is called “training” the algorithm. Once the algorithm has been trained,
the inferred pattern can be used to predict output values based on new inputs (i.e.,
ones not in the training data set).
Multiple regression is an example of supervised learning. A regression model
takes matched data (X’s, Y) and uses it to estimate parameters that characterize the
relationship between Y and the X’s. The estimated parameters can then be used to
predict Y on a new, different set of X’s. The difference between the predicted and
actual Y is used to evaluate how well the regression model predicts out-­of-­sample
(i.e., using new data).
The terminology used with ML algorithms differs from that used in regression.
Exhibit 1 provides a visual of the supervised learning model training process and a
translation between regression and ML terminologies.
 © CFA Institute. For candidate use only. Not for distribution.
466 Reading 7 ■ Machine Learning

Exhibit 1  Overview of Supervised Learning

Test dataset:
Regression ML Inputs only.
Test Data
Y = Output = Dependent variable = Target
{Xk k=1, N}
Xk = Inputs = Independent variables = Features

Labeled Data Supervised ML Prediction {Y, Xk k=1, N}

{Y, Xk k=1, N} Algorithm Rule Test Data

Training dataset:
Target outcomes Prediction Actual
paired with {YPredict} {Y}
feature inputs.

Evaluation of Fit

In supervised machine learning, the dependent variable (Y) is the target and the
independent variables (X’s) are known as features. The labeled data (training data set)
is used to train the supervised ML algorithm to infer a pattern-­based prediction rule.
The fit of the ML model is evaluated using labeled test data in which the predicted
targets (YPredict) are compared to the actual targets (YActual).
An example of supervised learning is the case in which ML algorithms are used
to predict whether credit card transactions are fraudulent or legitimate. In the credit
card example, the target is a binary variable with a value of 1 for “fraudulent” or 0
for “non-­fraudulent.” The features are the transaction characteristics. The chosen ML
algorithm uses these data to train a model to predict fraud more accurately in new
transactions. The ML program “learns from experience” if the percentage of correctly
predicted credit card transactions increases as the input from a growing credit card
database increases. One possible ML algorithm to use here would be to fit a logistic
regression model to the data to provide an estimate of the probability a transaction
is fraudulent.
Supervised learning can be divided into two categories of problems, regression
problems and classification problems, with the distinction between them being deter-
mined by the nature of the target (Y) variable. If the target variable is continuous, then
the task is one of regression (even if the ML technique used is not “regression,” note
this nuance of ML terminology). If the target variable is categorical or ordinal (i.e.,
a ranked category), then it is a classification problem. Regression and classification
use different ML techniques.
Regression focuses on making predictions of continuous target variables. Most
readers are already familiar with multiple linear regression (e.g., ordinary least squares)
models, but other supervised learning techniques exist that include non-­linear models.
These non-­linear models are useful for problems involving large data sets with large
numbers of features, many of which may be correlated. Some examples of problems
belonging to the regression category are using historical stock market returns to fore-
cast stock price performance or using historical corporate financial ratios to forecast
the probability of bond default.
Classification focuses on sorting observations into distinct categories. In a regres-
sion problem, when the dependent variable (target) is categorical, the model relating
the outcome to the independent variables (features) is called a “classifier.” Many
classification models are binary classifiers, as in the case of fraud detection for credit
card transactions. Multi-­category classification is not uncommon, as in the case of
classifying firms into multiple credit rating categories. In assigning ratings, the outcome
 © CFA Institute. For candidate use only. Not for distribution.
What Is Machine Learning? 467

variable is ordinal, meaning the categories have a distinct order or ranking (e.g., from
low to high creditworthiness). Ordinal variables are intermediate between categorical
variables and continuous variables on a scale of measurement.

3.3  Unsupervised Learning

Unsupervised learning is machine learning that does not make use of labeled data.
More formally, in unsupervised learning, we have inputs (X’s) that are used for anal-
ysis without any target (Y) being supplied. In unsupervised learning, because the ML
algorithm is not given labeled training data, the algorithm seeks to discover structure
within the data themselves. As such, unsupervised learning is useful for exploring
new data sets as it can provide human experts with insights into a data set too big or
too complex to visualize.
Two important types of problems that are well suited to unsupervised machine
learning are reducing the dimension of data and sorting data into clusters, known as
dimension reduction and clustering, respectively.
Dimension reduction focuses on reducing the number of features while retaining
variation across observations to preserve the information contained in that variation.
Dimension reduction may have several purposes. It may be applied to data with a large
number of features to produce a lower dimensional representation (i.e., with fewer
features) that can fit, for example, on a computer screen. Dimension reduction is also
used in many quantitative investment and risk management applications where it is
critical to identify the major factors underlying asset price movements.
Clustering focuses on sorting observations into groups (clusters) such that obser-
vations in the same cluster are more similar to each other than they are to observations
in other clusters. Groups are formed based on a set of criteria that may or may not
be prespecified (such as the number of groups). Clustering has been used by asset
managers to sort companies into empirically determined groupings (e.g., based on
their financial statement data) rather than conventional groupings (e.g., based on
sectors or countries).

3.4  Deep Learning and Reinforcement Learning

More broadly in the field of artificial intelligence, additional categories of machine
learning algorithms are distinguished. In deep learning, sophisticated algorithms
address highly complex tasks, such as image classification, face recognition, speech
recognition, and natural language processing. In reinforcement learning, a com-
puter learns from interacting with itself (or data generated by the same algorithm).
Deep learning and reinforcement learning are themselves based on neural networks
(NNs, also called artificial neural networks, or ANNs). NNs include highly flexible
ML algorithms that have been successfully applied to a variety of tasks characterized
by non-­linearities and interactions among features. Besides being commonly used
for classification and regression, neural networks are also the foundation for deep
learning and reinforcement learning, which can be either supervised or unsupervised.

3.5  Summary of ML Algorithms and How to Choose Among

Them
Exhibit 2 is a guide to the various machine learning algorithms organized by algorithm
type (supervised or unsupervised) and by type of variables (continuous, categorical, or
both). We will not cover linear or logistic regression since they are covered elsewhere
 © CFA Institute. For candidate use only. Not for distribution.
468 Reading 7 ■ Machine Learning

in quantitative methods. The extensions of linear regression, such as penalized regres-

sion and least absolute shrinkage and selection operator (LASSO), as well as the other
ML algorithms shown in Exhibit 2, will be covered in this reading.

Exhibit 2  Guide to ML Algorithms

ML Algorithm Type
Supervised Unsupervised
Variables (Target Variable) (No Target Variable)

Continuous Regression Dimensionality Reduction

• Linear; Penalized Regression/LASSO • Principal Components Analysis (PCA)
• Logistic Clustering
• Classification and Regression Tree (CART) • K-­Means
• Random Forest • Hierarchical
Categorical Classification Dimensionality Reduction
• Logit • Principal Components Analysis (PCA)
• Support Vector Machine (SVM) Clustering
• K-­Nearest Neighbor (KNN) • K-­Means
• Classification and Regression Tree (CART) • Hierarchical
Continuous or Neural Networks Neural Networks
Categorical Deep Learning Deep Learning
Reinforcement Learning Reinforcement Learning

Exhibit 3 presents a stylized decision flowchart for choosing among the machine
learning algorithms shown in Exhibit 2. The dark-­shaded ovals contain the supervised
ML algorithms; the light-­shaded ovals contain the unsupervised ML algorithms; and
the key questions to consider are shown in the unshaded boxes.
 © CFA Institute. For candidate use only. Not for distribution.
What Is Machine Learning? 469

Exhibit 3  Stylized Decision Flowchart for Choosing ML Algorithms

Classification

KNN or CART,
Random Forests, Dimensionality
SVM Start
Neural Nets Reduction

No Yes PCA
Complex Yes
Non-linear Complexity
Data? Reduction?

No
Yes
Labeled Yes No Numerical
Data? Classification Prediction?
No Yes
Complex
Non-linear Complex
No Data? Non-linear
Yes Data?
No Yes
Number of
Categories Neural
Known? Nets Penalized CART,
No
Yes Regression, Random Forests,
LASSO Neural Nets
Hierarchical
K-Means
Clustering
Regression
Clustering

First, start by asking, are the data complex, having many features that are highly
correlated? If yes, then dimensionality reduction using principal components analysis
(PCA) is appropriate.
Next, is the problem one of classification or numerical prediction? If numerical
prediction, then depending on whether or not the data have non-­linear characteristics,
the choice of ML algorithms is from a set of regression algorithms—either penalized
regression/LASSO for linear data or CART, random forest, and neural networks for
non-­linear data.
If the problem is one of classification, then depending on whether or not the data
are labeled, the choice is either from a set of classification algorithms using labeled
data or from a set of clustering algorithms using unlabeled data.
If the data are labeled, then depending on whether or not the data have non-­linear
characteristics, the choice of classification algorithm would be k-­nearest neighbor
(KNN) and support vector machine (SVM) for linear data or CART, random forest,
and neural networks for non-­linear data.
Finally, if the data are unlabeled, the choice of clustering algorithm depends on
whether or not the data have non-­linear characteristics. The choice of clustering
algorithm would be neural networks for non-­linear data or for linear data, k-­means
with a known number of categories, and hierarchical clustering with an unknown
number of categories.
Following a description of how to evaluate ML algorithm performance, we will
define all of the ML algorithms shown in Exhibit 3 and then explain their applications
in investment management.
 © CFA Institute. For candidate use only. Not for distribution.
470 Reading 7 ■ Machine Learning

EXAMPLE 1 

Machine Learning Overview

1 Which of the following best describes machine learning? Machine
learning:
A is a type of computer algorithm.
B is a set of computer-­driven approaches aimed at generating structure
or predictions from data without human intervention by finding a
pattern and then applying the pattern.
C is a set of computer-­driven approaches adapted to extracting informa-
tion from linear, labeled data sets.
2 Which of the following statements is most accurate? When attempting to
discover groupings of data without any target (Y) variable:
A an unsupervised ML algorithm is used.
B an ML algorithm that is given labeled training data is used.
C a supervised ML algorithm is used.
3 Which of the following statements concerning supervised learning best
distinguishes it from unsupervised learning? Supervised learning involves:
A training on labeled data to infer a pattern-­based prediction rule.
B training on unlabeled data to infer a pattern-­based prediction rule.
C learning from unlabeled data by discovering underlying structure in
the data themselves.
4 Which of the following best describes dimension reduction? Dimension
reduction:
A focuses on classifying observations in a data set into known groups
using labeled training data.
B focuses on clustering observations in a data set into unknown groups
using unlabeled data.
C focuses on reducing the number of features in a data set while retain-
ing variation across observations to preserve the information in that
variation.

Solution to 1:
B is correct. A is incorrect because machine learning is not best described as
a type of computer algorithm. C is incorrect because machine learning is not
limited to extracting information from linear, labeled data sets.
Solution to 2:
A is correct. B is incorrect because the term “labeled training data” means the
target (Y) is provided. C is incorrect because a supervised ML algorithm is meant
to predict a target (Y) variable.
Solution to 3:
A is correct. B is incorrect because supervised learning uses labeled training
data. C is incorrect because it describes unsupervised learning.
Solution to 4:
C is correct. A is incorrect because it describes classification, not dimension
reduction. B is incorrect because it describes clustering, not dimension reduction.
 © CFA Institute. For candidate use only. Not for distribution.
Overview of Evaluating ML Algorithm Performance 471

OVERVIEW OF EVALUATING ML ALGORITHM

PERFORMANCE 4
Machine learning algorithms promise several advantages relative to a structured
statistical approach in exploring and analyzing the structure of very large data sets.
ML algorithms have the ability to uncover complex interactions between feature vari-
ables and the target variable, and they can process massive amounts of data quickly.
Moreover, many ML algorithms can easily capture non-­linear relationships and may
be able to recognize and predict structural changes between features and the target.
These advantages mainly derive from the non-­parametric and non-­linear models that
allow more flexibility when inferring relationships.
The flexibility of ML algorithms comes with a price, however. ML algorithms
can produce overly complex models with results that are difficult to interpret, may
be sensitive to particular changes in the data, and may fit the training data too well.
An ML algorithm that fits the training data too well will typically not predict well
using new data. This problem is known as overfitting, and it means that the fitted
algorithm does not generalize well to new data. A model that generalizes well is a
model that retains its explanatory power when predicting out-­of-­sample (i.e., using
new data). An overfit model has incorporated the noise or random fluctuations in
the training data into its learned relationship. The problem is that these aspects do
not apply to any new data the algorithm receives and so will negatively impact the
model’s ability to generalize. The evaluation of any ML algorithm thus focuses on its
prediction error on new data rather than on its goodness-­of-­fit on the data in which
the algorithm was fitted (i.e., trained).
Generalization is an objective in model building, so the problem of overfitting
is a challenge to attaining that objective. These two concepts are the focus of the
discussion below.

4.1  Generalization and Overfitting

To properly describe generalization and overfitting of an ML model, it is important
to note the partitioning of the data set to which the model will be applied. The data
set is typically divided into three non-­overlapping samples: 1) training sample used
to train the model; 2) validation sample for validating and tuning the model; and 3)
test sample for testing the model’s ability to predict well on new data. The training
sample is often referred to as being “in-­sample,” while the validation and test samples
are commonly referred to as being “out-­of-­sample.” We will return shortly to the topic
of partitioning the data set.
To be valid and useful, any supervised machine learning model must generalize
well beyond the training data. The model should retain its explanatory power when
tested out-­of-­sample. As mentioned, one common reason for failure to generalize is
overfitting. Think of overfitting as tailoring a custom suit that fits only one person.
Continuing the analogy, underfitting is similar to making a baggy suit that fits no one,
whereas robust fitting, the desired result, is similar to fashioning a universal suit that
fits all similar people.
The concepts of underfitting, overfitting, and good (or robust) fitting are illustrated
in Exhibit 4. Underfitting means the model does not capture the relationships in the
data. The left graph shows four errors in this underfit model (three misclassified circles
and one misclassified triangle). Overfitting means training a model to such a degree of
specificity to the training data that the model begins to incorporate noise coming from
quirks or spurious correlations; it mistakes randomness for patterns and relationships.
The algorithm may have memorized the data, rather than learned from it, so it has
 © CFA Institute. For candidate use only. Not for distribution.
472 Reading 7 ■ Machine Learning

perfect hindsight but no foresight. The main contributors to overfitting are thus high
noise levels in the data and too much complexity in the model. The middle graph shows
no errors in this overfit model. Complexity refers to the number of features, terms, or
branches in the model and to whether the model is linear or non-­linear (non-­linear is
more complex). As models become more complex, overfitting risk increases. A good
fit/robust model fits the training (in-­sample) data well and generalizes well to out-­of-­
sample data, both within acceptable degrees of error. The right graph shows that the
good fitting model has only one error, the misclassified circle.

Exhibit 4  Under-, Over-, and Good-­Fitting

Y Y Y

Underfit X Overfit X Good Fit X

4.2  Errors and Overfitting

To capture these effects and calibrate degree of fit, data scientists compare error rates
in- and out-­of-­sample as a function of both the data and the algorithm. Total in-­
sample errors (Ein) are generated by the predictions of the fitted relationship relative
to actual target outcomes on the training sample. Total out-­of-­sample errors (Eout)
are from either the validation or test samples. Low or no in-­sample error but large
out-­of-­sample error are indicative of poor generalization. Data scientists decompose
the total out-­of-­sample error into three sources:
1 Bias error, or the degree to which a model fits the training data. Algorithms
with erroneous assumptions produce high bias with poor approximation, caus-
ing underfitting and high in-­sample error.
2 Variance error, or how much the model’s results change in response to new
data from validation and test samples. Unstable models pick up noise and pro-
duce high variance, causing overfitting and high out-­of-­sample error.
3 Base error due to randomness in the data.
A learning curve plots the accuracy rate (= 1 – error rate) in the validation or
test samples (i.e., out-­of-­sample) against the amount of data in the training sample,
so it is useful for describing under- and overfitting as a function of bias and variance
errors. If the model is robust, out-­of-­sample accuracy increases as the training sample
size increases. This implies that error rates experienced in the validation or test sam-
ples (Eout) and in the training (Ein) sample converge toward each other and toward a
desired error rate (or, alternatively, the base error). In an underfitted model with high
bias error, shown in the left panel of Exhibit 5, high error rates cause convergence
below the desired accuracy rate. Adding more training samples will not improve the
model. In an overfitted model with high variance error, shown in the middle panel of
Exhibit 5, the validation sample and training sample error rates fail to converge. In
 © CFA Institute. For candidate use only. Not for distribution.
Overview of Evaluating ML Algorithm Performance 473

building models, data scientists try to simultaneously minimize both bias and variance
errors while selecting an algorithm with good predictive or classifying power, as seen
in the right panel of Exhibit 5.

Exhibit 5  Learning Curves: Accuracy in Validation and Training Samples vs.

Training Sample Size
C. Good Tradeoff of Bias
A. High Bias Error B. High Variance Error and Variance Error
Accuracy Rate (%) Accuracy Rate (%) Accuracy Rate (%)
100 100 100

0 0 0
Number of Training Samples Number of Training Samples Number of Training Samples

Desired Accuracy Rate Training Accuracy Rate

Validation Accuracy Rate

Out-­of-­sample error rates are also a function of model complexity. As complexity

increases in the training set, error rates (Ein) fall and bias error shrinks. As complex-
ity increases in the test set, however, error rates (Eout) rise and variance error rises.
Typically, linear functions are more susceptible to bias error and underfitting, while
non-­linear functions are more prone to variance error and overfitting. Therefore, an
optimal point of model complexity exists where the bias and variance error curves
intersect and in- and out-­of-­sample error rates are minimized. A fitting curve, which
shows in- and out-­of-­sample error rates (Ein and Eout) on the y-­axis plotted against
model complexity on the x-­axis, is presented in Exhibit 6 and illustrates this trade-­off.

Exhibit 6  Fitting Curve Shows Trade-­Off Between Bias and Variance Errors
and Model Complexity
Model Error (Ein, Eout)

Optimal
Complexity

Total
Error

Variance Bias
Error Error

Model Complexity
 © CFA Institute. For candidate use only. Not for distribution.
474 Reading 7 ■ Machine Learning

Finding the optimal point (managing overfitting risk)—the sweet spot just before
the total error rate starts to rise (due to increasing variance error)—is a core part of
the machine learning process and the key to successful generalization. Data scientists
express the trade-­off between overfitting and generalization as a trade-­off between
cost (the difference between in- and out-­of-­sample error rates) and complexity. They
use the trade-­off between cost and complexity to calibrate and visualize under- and
overfitting and to optimize their models.

4.3  Preventing Overfitting in Supervised Machine Learning

We have seen that overfitting impairs generalization, but overfitting potential is endemic
to the supervised machine learning process due to the presence of noise. So, how do
data scientists combat this risk? Two common guiding principles and two methods
are used to reduce overfitting: 1) preventing the algorithm from getting too complex
during selection and training, which requires estimating an overfitting penalty; and 2)
proper data sampling achieved by using cross-­validation, a technique for estimating
out-­of-­sample error directly by determining the error in validation samples.
The first strategy comes from “Occam’s razor,” the problem-­solving principle that
the simplest solution tends to be the correct one. In supervised machine learning, it
means limiting the number of features and penalizing algorithms that are too complex
or too flexible by constraining them to include only parameters that reduce out-­of-­
sample error.
The second strategy comes from the principle of avoiding sampling bias. Good
experimental designs have representative samples and control groups, and both should
be drawn from the same theoretical distribution. But sampling bias can creep into
machine learning in many ways. The challenge is having a large enough data set to
make both training and testing possible on representative samples. An unrepresen-
tative sample or reducing the training sample size too much could obscure its true
patterns, thereby increasing bias. In supervised machine learning, the technique
for reducing sampling bias is through careful partitioning of the data set into three
groups: 1) training sample, the set of labeled training data where the target variable
(Y) is known; 2) validation sample, the set of data for comparing various solutions and
tuning the selected model, thereby validating the model; and 3) test sample, the set of
data held aside for testing to confirm the model’s predictive or classifying power. The
goal, of course, is to deploy the tested model on fresh data from the same domain.
To mitigate the problem of such holdout samples (i.e., data samples not used to
train the model) reducing the training set size too much, modelers use special cross-­
validation techniques. One such technique is k-­fold cross-­validation, in which the
data (excluding test sample and fresh data) are shuffled randomly and then are divided
into k equal sub-­samples, with k – 1 samples used as training samples and one sample,
the kth, used as a validation sample. Note that k is typically set at 5 or 10. This process
is then repeated k times, which helps minimize both bias and variance by insuring
that each data point is used in the training set k – 1 times and in the validation set
once. The average of the k validation errors (mean Eval) is then taken as a reasonable
estimate of the model’s out-­of-­sample error (Eout).
In sum, mitigating overfitting risk by avoiding excessive out-­of-­sample error is
critical to creating a supervised machine learning model that generalizes well to
fresh data sets drawn from the same domain. The main techniques used to mitigate
overfitting risk in model construction are complexity reduction and cross-­validation.
 © CFA Institute. For candidate use only. Not for distribution.
Overview of Evaluating ML Algorithm Performance 475

EXAMPLE 2 

Evaluating ML Algorithm Performance

Shreya Anand is a portfolio manager based in the Mumbai headquarters office
of an investment firm, where she runs a high-­dividend yield fund for wealthy
clients. Anand has some knowledge of data science from her course in univer-
sity. She is interested in classifying companies in the NIFTY 200 Index—an
index of large- and mid-­cap companies listed on the National Stock Exchange
of India—into two categories: dividend increase and no dividend increase. She
assembles data for training, validating, and testing an ML-­based model that
consists of 1,000 observations of NIFTY 200 companies, each consisting of 25
features (fundamental and technical) and the labeled target (dividend increase
or no dividend increase).
After training her model, Anand discovers that while it is good at correctly
classifying using the training sample, it does not perform well using new data. In
consulting her colleagues about this issue, Anand hears conflicting explanations
about what constitutes good generalization in an ML model:
Statement 1 The model retains its explanatory power when predicting
using new data (i.e., out-­of-­sample).
Statement 2 The model shows low explanatory power after training
using in-­sample data (i.e., training data).
Statement 3 The model loses its explanatory power when predicting
using new data (i.e., out-­of-­sample).

1 Which statement made to Anand is most accurate?

A Statement 1.
B Statement 2.
C Statement 3.
2 Anand’s model is most likely being impaired by which of the following:
A Underfitting and bias error
B Overfitting and variance error.
C Overfitting and bias error.
3 By implementing which one of the following actions can Anand address
the problem?
A Estimate and incorporate into the model a penalty that decreases in
size with the number of included features.
B Use the k-­fold cross-­validation technique to estimate the model’s out-­
of-­sample error, and then adjust the model accordingly.
C Use an unsupervised learning model.

Solution to 1:
A, Statement 1, is correct. B, Statement 2, is incorrect because it describes
a poorly fitting model with high bias. C, Statement 3, is incorrect because it
describes an overfitted model with poor generalization.
Solution to 2:
B is correct. Anand’s model is good at correctly classifying using the training
sample, but it does not perform well using new data. The model is overfitted,
so it has high variance error.
 © CFA Institute. For candidate use only. Not for distribution.
476 Reading 7 ■ Machine Learning

Solution to 3:
B is correct. A is incorrect because the penalty should increase in size with the
number of included features. C is incorrect because Anand is using labeled data
for classification, and unsupervised learning models do not use labeled data.

5 SUPERVISED MACHINE LEARNING ALGORITHMS

Supervised machine learning models are trained using labeled data, and depending
on the nature of the target (Y) variable, they can be divided into two types: regression
for a continuous target variable and classification for a categorical or ordinal target
variable. As shown in Exhibits 2 and 3 under regression, we will now cover penalized
regression and LASSO. Then as shown under classification, we will introduce support
vector machine (SVM), k-­nearest neighbor (KNN), and classification and regression
tree (CART) algorithms. Note that CART, as its name implies, can be used for both
classification and regression problems.
In the following discussion, assume we have a number of observations of a target
variable, Y, and n real-­valued features, X1, . . . , Xn, that we may use to establish a
relationship (regression or classification) between X (a vector of the Xi) and Y for each
observation in our data set.

5.1  Penalized Regression

Penalized regression is a computationally efficient technique used in prediction prob-
lems. In practice, penalized regression has been useful for reducing a large number
of features to a manageable set and for making good predictions in a variety of large
data sets, especially where features are correlated (i.e., when classical linear regression
breaks down).
In a large data set context, we may have many features that potentially could be
used to explain Y. When a model is fit to training data, the model may so closely reflect
the characteristics of the specific training data that the model does not perform well
on new data. Features may be included that reflect what is noise or randomness in the
training data set that will not be present in new or future data used for making pre-
dictions. That is the problem of overfitting, and penalized regression can be described
as a technique to avoid overfitting. In prediction, out-­of-­sample performance is key,
so relatively parsimonious models (that is, models in which each variable plays an
essential role) tend to work well because they are less subject to overfitting.
Let us suppose that we standardize our data, so the features have a mean of 0 and
a variance of 1. Standardization of features will allow us to compare the magnitudes
of regression coefficients for the feature variables. In ordinary linear regression (i.e.,
ordinary least squares, or OLS), the regression coefficients b 0 , b1, . . . , b K are chosen
to minimize the sum of the squared residuals (i.e., the sum of the squared difference
between the actual values, Yi, and the predicted values, Y i ), or:
n
∑(Yi − Y i )
2

i =1
Penalized regression includes a constraint such that the regression coefficients
are chosen to minimize the sum of squared residuals plus a penalty term that increases
in size with the number of included features. So, in a penalized regression, a feature
 © CFA Institute. For candidate use only. Not for distribution.
Supervised Machine Learning Algorithms 477

must make a sufficient contribution to model fit to offset the penalty from including
it. Therefore, only the more important features for explaining Y will remain in the
penalized regression model.
In one popular type of penalized regression, LASSO (least absolute shrinkage and
selection operator), the penalty term has the following form, with λ > 0:
K
Penalty term = λ ∑ b k .
k =1
In addition to minimizing the sum of the squared residuals, LASSO also involves
minimizing the sum of the absolute values of the regression coefficients (see the fol-
lowing expression). The greater the number of included features (i.e., variables with
non-­zero coefficients), the larger the penalty term. Therefore, penalized regression
ensures that a feature is included only if the sum of squared residuals declines by more
than the penalty term increases. All types of penalized regression involve a trade-­off of
this type. Also, since LASSO eliminates the least important features from the model,
it automatically performs feature selection.
n K
2
∑(Yi − Yi ) + λ ∑ b K
i =1 k =1
Lambda (λ) is a hyperparameter—a parameter whose value must be set by the
researcher before learning begins—of the regression model and will determine the
balance between fitting the model versus keeping the model parsimonious. Note that
in the case where λ = 0, then the LASSO penalized regression is equivalent to an
OLS regression. When using LASSO or other penalized regression techniques, the
penalty term is added only during the model building process (i.e., when fitting the
model to the training data). Once the model has been built, the penalty term is no
longer needed, and the model is then evaluated by the sum of the squared residuals
generated using the test data set.
With today’s availability of fast computation algorithms, investment analysts are
increasingly using LASSO and other regularization techniques to remove less per-
tinent features and build parsimonious models. Regularization describes methods
that reduce statistical variability in high dimensional data estimation problems—in
this case, reducing regression coefficient estimates toward zero and thereby avoiding
complex models and the risk of overfitting. LASSO has been used, for example, for
forecasting default probabilities in industrial sectors where scores of potential features,
many collinear, have been reduced to fewer than 10 variables, which is important given
the relatively small number (about 100) of observations of default.
Regularization methods can also be applied to non-­linear models. A long-­term
challenge of the asset management industry in applying mean–variance optimization
has been the estimation of stable covariance matrixes and asset weights for large
portfolios. Asset returns typically exhibit strong multi-­collinearity, making the esti-
mation of the covariance matrix highly sensitive to noise and outliers, so the resulting
optimized asset weights are highly unstable. Regularization methods have been used
to address this problem.
In prediction, only out-­of-­sample performance (i.e., prediction accuracy) really
matters. The relatively parsimonious models produced by applying penalized regression
methods, like LASSO, tend to work well because they are less subject to overfitting.

5.2  Support Vector Machine

Support vector machine (SVM) is one of the most popular algorithms in machine
learning. It is a powerful supervised algorithm used for classification, regression,
and outlier detection. Despite its complicated sounding name, the notion is relatively
 © CFA Institute. For candidate use only. Not for distribution.
478 Reading 7 ■ Machine Learning

straightforward and best explained with a few pictures. The left panel in Exhibit 7
presents a simple data set with two features (x and y coordinates) labeled in two groups
(triangles and crosses). These binary labeled data are noticeably separated into two
distinct regions, which could represent stocks with positive and negative returns in
a given year. These two regions can be easily separated by numerous straight lines;
three of them are shown in the right panel of Exhibit 7. The data are thus linearly
separable, and any of the straight lines shown would be called a linear classifier—a
binary classifier that makes its classification decision based on a linear combination
of the features of each data point.

Exhibit 7:  Scatterplots and Linear Separation of Labeled Data

A. Data Labeled in Two Groups B. Data is Linearly Separable

Y Y

X X

With two dimensions or features (x and y) linear classifiers can be represented

as straight lines. Observations with n features can be represented in an n-dimension
space, and the data set would be linearly separable if the observations can be sepa-
rated into two distinct regions by a linear space boundary. The general term for such
a space boundary is an n-dimensional hyperplane, which with n = 1 is called a line
and with n = 2 is called a plane.
Support vector machine (SVM) is a linear classifier that determines the hyper-
plane that optimally separates the observations into two sets of data points. The
intuitive idea behind the SVM algorithm is maximizing the probability of making a
correct prediction (here, that an observation is a triangle or a cross) by determining
the boundary that is the furthest away from all the observations. In Exhibit 8, SVM
separates the data by the maximum margin, where the margin is the shaded strip that
divides the observations into two groups. The straight line in the middle of the shaded
strip is the discriminant boundary or boundary, for short. We can see that the SVM
algorithm produces the widest shaded strip (i.e., the one with the maximum margin
on either side of the boundary). The margin is determined by the observations closest
to the boundary (the circled points) in each set, and these observations are called
support vectors. Adding more training data away from the support vectors will not
affect the boundary. In our training data sets, however, the result will be sensitive to
the support vectors.
 © CFA Institute. For candidate use only. Not for distribution.
Supervised Machine Learning Algorithms 479

Exhibit 8  Linear Support Vector Machine Classifier

Y

Margin

In Exhibit 8, SVM is classifying all observations perfectly. Many real-­world data

sets, however, are not perfectly linearly separable. Some observations may fall on the
wrong side of the boundary and be misclassified by the SVM algorithm. The SVM
algorithm handles this problem by an adaptation called soft margin classification,
which adds a penalty to the objective function for observations in the training set that
are misclassified. In essence, the SVM algorithm will choose a discriminant boundary
that optimizes the trade-­off between a wider margin and a lower total error penalty.
As an alternative to soft margin classification, a non-­linear SVM algorithm can
be run by introducing more advanced, non-­linear separation boundaries. These algo-
rithms will reduce the number of misclassified instances in the training data sets but
will have more features, thus adding to the model’s complexity.
SVM has many applications in investment management. It is particularly suited
for small- to medium-­size but complex high-­dimensional data sets, such as corporate
financial statements or bankruptcy databases. Investors seek to predict company fail-
ures for identifying stocks to avoid or to short sell, and SVM can generate a binary
classification (e.g., bankruptcy likely vs. bankruptcy unlikely) using many fundamental
and technical feature variables. SVM can quickly capture the characteristics of such
data with many features while being resilient to outliers and correlated features.
SVM can also be used to classify text from documents (e.g., news articles, company
announcements, and company annual reports) into useful categories for investors
(e.g., positive sentiment and negative sentiment).

5.3  K-­Nearest Neighbor

K-­nearest neighbor (KNN) is a supervised learning technique most often used for
classification and sometimes for regression. The idea is to classify a new observation
by finding similarities (“nearness”) between this new observation and the existing data.
Going back to the scatterplot in Exhibit 7, let us assume we have a new observation:
The diamond in Exhibit 9 needs to be classified as belonging to either the cross or the
triangle category. If k = 1, the diamond will be classified into the same category as its
nearest neighbor (i.e., the triangle in the left panel). The right panel in Exhibit 9 presents
the case where k = 5, so the algorithm will look at the diamond’s 5 nearest neighbors,
which are 3 triangles and 2 crosses. The decision rule is to choose the classification
with the largest number of nearest neighbors out of the 5 being considered. So, the
diamond is again classified as belonging to the triangle category.
 © CFA Institute. For candidate use only. Not for distribution.
480 Reading 7 ■ Machine Learning

Exhibit 9   K-­Nearest Neighbor (KNN) Algorithm

A. KNN With New Observation, K=1 B. KNN With New Observation, K=5

Y Y

X X

Let us suppose we have a database of corporate bonds classified by credit rating

that also contains detailed information on the characteristics of these bonds. Such
features would include those of the issuing company (e.g., asset size, industry, leverage
ratios, cash flow ratios) and of the bond issue itself (e.g., tenor, fixed/floating coupon,
embedded options). Now, assume a new bond is about to be issued with no credit
rating. By nature, corporate bonds with similar issuer and issue characteristics should
be given a similar credit rating. So, by using KNN, we can predict the implied credit
rating of the new bond based on the similarities of its characteristics to those of the
bonds in our database.
KNN is a straightforward, intuitive model that is still very powerful because it is
non-­parametric; the model makes no assumptions about the distribution of the data.
Moreover, it can be used directly for multi-­class classification. A critical challenge of
KNN, however, is defining what it means to be “similar” (or near). Besides the selec-
tion of features, an important decision relates to the distance metric used to model
similarity because an inappropriate measure will generate poorly performing models.
The choice of a correct distance measure may be even more subjective for ordinal
or categorical data. For example, if an analyst is looking at the similarities in market
performance of various equities, he/she may consider using the correlation between
the stocks’ historical returns as an appropriate measure of similarity.
Knowledge of the data and understanding of the business objectives of the analysis
are critical steps in the process of defining similarity. KNN results can be sensitive to
inclusion of irrelevant or correlated features, so it may be necessary to select features
manually. By doing so, the analyst removes less-­valuable information to keep the most
relevant and pertinent information. If done correctly, this process should generate a
more representative distance measure. KNN algorithms tend to work better with a
small number of features.
Finally, the number k, the hyperparameter of the model, must be chosen with the
understanding that different values of k can lead to different conclusions. For predicting
the credit rating of an unrated bond, for example, should k be the 3, 15, or 50 most
similar bonds to the unrated bond? If k is an even number, there may be ties and no
clear classification. Choosing a value for k that is too small would result in a high error
rate and sensitivity to local outliers, but choosing a value for k that is too large would
dilute the concept of nearest neighbors by averaging too many outcomes. In practice,
several different techniques can be used to determine an optimal value for k, taking
into account the number of categories and their partitioning of the feature space.
The KNN algorithm has many applications in the investment industry, including
bankruptcy prediction, stock price prediction, corporate bond credit rating assignment,
and customized equity and bond index creation.
 © CFA Institute. For candidate use only. Not for distribution.
Supervised Machine Learning Algorithms 481

5.4  Classification and Regression Tree

Classification and Regression Tree (CART) is another common supervised machine
learning technique that can be applied to predict either a categorical target variable,
producing a classification tree, or a continuous target variable, producing a regression
tree. Most commonly, CART is applied to binary classification or regression.
CART will be discussed in the context of a simplified model for classifying com-
panies by whether or not they increase their dividends to shareholders. Such a classi-
fication requires a binary tree: a combination of an initial root node, decision nodes,
and terminal nodes. The root node and each decision node represent a single feature
(f) and a cutoff value (c) for that feature. As shown in Panel A of Exhibit 10, we start at
the initial root node for a new data point. In this case, the initial root node represents
the feature investment opportunities growth (IOG), designated as X1, with a cutoff
value of 10%. From the initial root node, the data are partitioned at decision nodes
into smaller and smaller subgroups until terminal nodes are formed that contain the
predicted labels. In this case, the predicted labels are either dividend increase (the
cross) or no dividend increase (the dash).
Also shown in Panel A of Exhibit 10, if the value of feature IOG (X1) is greater than
10% (Yes), then we proceed to the decision node for free cash flow growth (FCFG),
designated as X2, which has a cutoff value of 20%. Now, if the value of FCFG is not
greater than 20% (No), then CART will predict that data point belongs to the no div-
idend increase (dash) category, which represents a terminal node. Conversely, if the
value of X2 is greater than 20% (Yes), then CART will predict that data point belongs
to the dividend increase (cross) category, which represents another terminal node.
It is important to note that the same feature can appear several times in a tree
in combination with other features. Moreover, some features may be relevant only
if other conditions have been met. For example, going back to the initial root node,
if IOG is not greater than 10% (X1 ≤10%) and FCFG is greater than 10%, then IOG
appears again as another decision node, but this time it is lower down in the tree and
has a cutoff value of 5%.
 © CFA Institute. For candidate use only. Not for distribution.
482 Reading 7 ■ Machine Learning

Exhibit 10  Classification and Regression Tree—Decision Tree and

Partitioning of the Feature Space
A. Decision Tree

IOG
X1>10%

No Yes

FCFG FCFG
X2>10% X2>20%

No Yes
Yes No

No Div.
Incr. Div.
IOG No Div.
(–) Incr.
X1>5% Incr. (+)
(–)

No Yes

Div. No Div.
Incr. Incr.
(+) (–)

X1: Investment Opportunities Growth (IOG)

X2: Free Cashflow Growth (FCFG)

B. Partitioning of the Feature (X1,X2) Feature Space

X1≤5% X1>5%
+ +
+ – +
X2>10% + + + X2>20%
–
+
– + +
– –
X2: FCFG –
X2≤20%
–
– –
X2≤10% –

X1≤10% X1>10%
X1: IOG

We now turn to how the CART algorithm selects features and cutoff values for
them. Initially, the classification model is trained from the labeled data, which in
this hypothetical case are 10 instances of companies having dividend increase (the
crosses) and 10 instances of companies with no dividend increase (the dashes). As
shown in Panel B of Exhibit 10, at the initial root node and at each decision node the
feature space (i.e., the plane defined by X1 and X2) is split into two rectangles for
values above and below the cutoff value for the particular feature represented at that
 © CFA Institute. For candidate use only. Not for distribution.
Supervised Machine Learning Algorithms 483

node. This can be seen by noting the distinct patterns of the lines that emanate from
the decision nodes in Panel A. These same distinct patterns are used for partitioning
the feature space in Panel B.
The CART algorithm chooses the feature and the cutoff value at each node that
generates the widest separation of the labeled data to minimize classification error
(e.g., by a criterion, such as mean-­squared error). After each decision node, the
partition of the feature space becomes smaller and smaller, so observations in each
group have lower within-­group error than before. At any level of the tree, when the
classification error does not diminish much more from another split (bifurcation),
the process stops, the node is a terminal node, and the category that is in the major-
ity at that node is assigned to it. If the objective of the model is classification, then
the prediction of the algorithm at each terminal node will be the category with the
majority of data points. For example, in Panel B of Exhibit 10, the top right rectangle
of the feature space, representing IOG (X1)>10% and FCFG (X2)>20%, contains 5
crosses, the most data points of any of the partitions. So, CART would predict that
a new data point (i.e., a company) with such features belongs to the cross (dividend
increase) category. However, if instead the new data point had IOG (X1)>10% and
FCFG (X2) ≤20%, then it would be predicted to belong to the dash (no dividend
increase) category—represented by the lower right rectangle with 2 crosses but with
3 dashes. Finally, if the goal is regression, then the prediction at each terminal node
is the mean of the labeled values.
CART makes no assumptions about the characteristics of the training data, so if
left unconstrained, potentially it can perfectly learn the training data. To avoid such
overfitting, regularization parameters can be added, such as the maximum depth of the
tree, the minimum population at a node, or the maximum number of decision nodes.
The iterative process of building nodes is stopped once the regularization criterion
has been reached. For example, in Panel B of Exhibit 10, the upper left rectangle of
the feature space (determined by X1≤10%, X2>10%, and X1≤5% with three crosses)
might represent a terminal node resulting from a regularization criterion with min-
imum population equal to 3. Alternatively, regularization can occur via a pruning
technique that can be used afterward to reduce the size of the tree. Sections of the
tree that provide little classifying power are pruned (i.e., removed).
By its iterative structure, CART can uncover complex dependencies between
features that other models cannot reveal. As demonstrated in Exhibit 10, the same
feature can appear several times in combination with other features and some features
may be relevant only if other conditions have been met.
As shown in Exhibit 11, high profitability is a critical feature for predicting if a
stock is an attractive investment or a value trap (i.e., an investment that, although
apparently priced cheaply, is likely to be unprofitable). This feature is relevant only
if the stock is cheap—for example, in this hypothetical case if P/E is less than 15,
leverage is high (debt to total capital > 50%), and sales are expanding (sales growth
> 15%). Said another way, high profitability is irrelevant in this context if the stock is
not cheap, and if leverage is not high, and if sales are not expanding. Multiple linear
regression typically fails in such situations where the relationship between the features
and the outcome is non-­linear.
 © CFA Institute. For candidate use only. Not for distribution.
484 Reading 7 ■ Machine Learning

Exhibit 11  Stylized Decision Tree—Attractive Investment or Value Trap?

Is Stock Cheap?
(P/E < 15x)

Yes No

Is Leverage High?
(Debt/Tot. Cap > 50%)

Yes No

Are Sales Growing?

(Sales Growth > 15%)

Yes No

Is Profitability High?
(Net Profit Margin > 20%)

Yes No

Attractive Equity
Value Trap
Investment

CART is a popular supervised machine learning model because the tree provides
a visual explanation for the prediction. This contrasts favorably with other algorithms
that are often considered to be “black boxes” because it may be difficult to understand
the reasoning behind their outcomes and thus to foster trust in them. CART is a
powerful tool to build expert systems for decision-­making processes. It can induce
robust rules despite noisy data and complex relationships between high numbers of
features. Typical applications of CART in investment management include, among
others, enhancing detection of fraud in financial statements, generating consistent
decision processes in equity and fixed-­income selection, and simplifying communi-
cation of investment strategies to clients.

5.5  Ensemble Learning and Random Forest

Instead of basing predictions on the results of a single model as in the previous dis-
cussion, why not use the predictions of a group—or an ensemble—of models? Each
single model will have a certain error rate and will make noisy predictions. But by
taking the average result of many predictions from many models, we can expect to
achieve a reduction in noise as the average result converges toward a more accurate
prediction. This technique of combining the predictions from a collection of models
is called ensemble learning, and the combination of multiple learning algorithms is
known as the ensemble method. Ensemble learning typically produces more accurate
and more stable predictions than the best single model. In fact, in many prestigious
machine learning competitions, an ensemble method is often the winning solution.
Ensemble learning can be divided into two main categories: 1) an ensemble method
can be an aggregation of heterogeneous learners (i.e., different types of algorithms
combined together with a voting classifier); or 2) an ensemble method can be an
 © CFA Institute. For candidate use only. Not for distribution.
Supervised Machine Learning Algorithms 485

aggregation of homogenous learners (i.e., a combination of the same algorithm, using

different training data that are based, for example, on a bootstrap aggregating (i.e.,
bagging) technique as discussed later).

5.5.1  Voting Classifiers

Suppose you have been working on a machine learning project for some time and
have trained and compared the results of several algorithms, such as SVM, KNN, and
CART. A majority-­vote classifier will assign to a new data point the predicted label
with the most votes. For example, if the SVM and KNN models are both predicting
the category “stock outperformance” and the CART model is predicting the category
“stock underperformance,” then the majority-­vote classifier will choose ‘“stock out-
performance.” The more individual models you have trained, the higher the accuracy
of the aggregated prediction up to a point. There is an optimal number of models
beyond which performance would be expected to deteriorate from overfitting. The
trick is to look for diversity in the choice of algorithms, modelling techniques, and
hypotheses. The (extreme) assumption here is that if the predictions of the individual
models are independent, then we can use the law of large numbers to achieve a more
accurate prediction.

5.5.2  Bootstrap Aggregating (Bagging)

Alternatively, one can use the same machine learning algorithm but with different
training data. Bootstrap aggregating (or bagging) is a technique whereby the
original training data set is used to generate n new training data sets or bags of data.
Each new bag of data is generated by random sampling with replacement from the
initial training set. The algorithm can now be trained on n independent data sets that
will generate n new models. Then, for each new observation, we can aggregate the
n predictions using a majority-­vote classifier for a classification or an average for a
regression. Bagging is a very useful technique because it helps to improve the stability
of predictions and protects against overfitting the model.

5.5.3  Random Forest

A random forest classifier is a collection of a large number of decision trees trained
via a bagging method. For example, a CART algorithm would be trained using each
of the n independent data sets (from the bagging process) to generate the multitude
of different decision trees that make up the random forest classifier.
To derive even more individual predictions, added diversity can be generated in
the trees by randomly reducing the number of features available during training. So,
if each observation has n features, one can randomly select a subset of m features
(where m < n) that will then be considered by the CART algorithm for splitting the
data set at each of the decision nodes. The number of subset features (m), the num-
ber of trees to use, the minimum size (population) of each node (or leaf ), and the
maximum depth of each tree are all hyperparameters that can be tuned to improve
overall model prediction accuracy. For any new observation, we let all the classifier
trees (the “random forest”) undertake classification by majority vote—implementing a
machine learning version of the “wisdom of crowds.” The process involved in random
forest tends to protect against overfitting on the training data. It also reduces the ratio
of noise to signal because errors cancel out across the collection of slightly different
classification trees. However, an important drawback of random forest is that it lacks
the ease of interpretability of individual trees; as a result, it is considered a relatively
black box-­type algorithm.
Exhibit 12 presents three scatterplots of actual and predicted defaults by small- and
medium-­sized businesses with respect to two features, X and Y—for example, firm
profitability and leverage, respectively. The left plot shows the actual cases of default
 © CFA Institute. For candidate use only. Not for distribution.
486 Reading 7 ■ Machine Learning

in light shade and no default in dark shade, while the middle and right plots present
the predicted defaults and no defaults (also in light and dark shades, respectively). It is
clear from the middle plot, which is based on a traditional linear regression model, that
the model fails to predict the complex non-­linear relationship between the features.
Conversely, the right plot, which presents the prediction results of a random forest
model, shows that this model performs very well in matching the actual distribution
of the data.

Exhibit 12  Credit Defaults of Small- and Medium-­Sized Borrowers

Actual Data Linear Model Prediction Random Forest Prediction

4 4 4

2 2 2

Y
0 0 0

–2 –2 –2

–2 0 2 –2 0 2 –2 0 2

X X X

Source: Bacham and Zhao (2017).

Ensemble Learning with Random Forest

In making use of voting across classifier trees, random forest is an example of ensemble
learning: Incorporating the output of a collection of models produces classifications that
have better signal-­to-­noise ratios than the individual classifiers. A good example of a
credit card fraud detection problem comes from an open source data set on Kaggle.1
Here, the data contained several anonymized features that might be used to explain
which transactions were fraudulent. The difficulty in the analysis arises from the fact that
the rate of fraudulent transactions is very low; in a sample of 284,807 transactions, only
492 were fraudulent (0.17%). This is akin to finding a needle in a haystack. Applying a
random forest classification algorithm with an oversampling technique—which involves
increasing the proportional representation of fraudulent data in the training set—does
extremely well. Despite the lopsided sample, it delivers precision (the percentage of cases
marked fraudulent that were actually fraudulent) of 89% and recall (the percentage of
fraudulent cases that were marked as fraudulent) of 82%.

Despite its relative simplicity, random forest is a powerful algorithm with many
investment applications. These include, for example, use in factor-­based investment
strategies for asset allocation and investment selection or use in predicting whether
an IPO will be successful (e.g., percent oversubscribed, first trading day close/IPO
price) given the attributes of the IPO offering and the corporate issuer.

1  See https://www.kaggle.com/mlg-­ulb/creditcardfraud (accessed 1 October 2018).

 © CFA Institute. For candidate use only. Not for distribution.
Supervised Machine Learning Algorithms 487

EXAMPLE 3 

Support Vector Machine and K-­Nearest Neighbor

Rachel Lee is a fixed-­income portfolio manager analyst with Zeta Investment
Management Company. Zeta manages an investment-­grade bond portfolio for
small, conservative institutions and a non-­investment-­grade (i.e., high-­yield)
bond portfolio for yield-­seeking, high-­net-­worth individuals. Both portfolios
can hold unrated bonds if the characteristics of the unrated bonds closely match
those of the respective portfolio’s average holding.
Lee is discussing an upcoming straight, 10-­year fixed coupon bond issue
with senior credit analyst Marc Watson. Watson comments that although the
bond’s issuer, Biotron Corporation, has not had this issue rated, his analysis of
the company’s profitability, cash flow, leverage, and coverage ratios places the
issue near the borderline between low investment-­grade (Baa3/BBB–) and high
non-­investment-­grade (Ba1/BB+) bonds.
Lee decides to use machine learning methods to confirm the implied credit
rating of Biotron Corporation.
1 State the type of problem being addressed by Lee.
2 State two ML algorithms that Lee could use to explore the implied credit
rating of Biotron Corporation, and then describe how each algorithm
could be applied.
Lee decides to apply the two identified ML algorithms. Both algorithms clearly
support a high non-­investment-­grade rating. Watson states that because both
ML algorithms agree on the rating, he has confidence in relying on the rating.
3 State one argument in support of Watson’s viewpoint.

Solution to 1:
Lee is addressing a supervised learning classification problem because she
must determine whether Biotron’s upcoming bond issue would be classified as
investment grade or non-­investment grade.
Solution to 2:
One suitable ML algorithm is the SVM. The SVM algorithm is a linear classifier
that aims to seek the optimal hyperplane—the one that separates observations
into two distinct sets by the maximum margin. So, the SVM is well suited to
binary classification problems, such as the one facing Lee (investment grade
vs. non-­investment grade). In this case, Lee could train the SVM algorithm on
data—characteristics (features) and rating (target)—of low investment-­grade
(Baa3/BBB–) and high non-­investment-­grade (Ba1/BB+) bonds. Lee would then
note on which side of the margin the new data point (Biotron’s new bonds) lies.
The KNN algorithm is also well suited for classification problems because
it classifies a new observation by finding similarities (or nearness) between the
new observation and the existing data. Training the algorithm with data as for
SVM, the decision rule for classifying Biotron’s new bonds is which classifi-
cation is in the majority among its k-nearest neighbors. Note that k must be
pre-­specified by Lee.
 © CFA Institute. For candidate use only. Not for distribution.
488 Reading 7 ■ Machine Learning

Solution to 3:
If the ML algorithms disagreed on the classification, the classification would
be more likely to be sensitive to the algorithm’s approach to classifying data.
Because the classification of Biotron’s new issue appears robust to the choice
of ML algorithm (i.e., both algorithms agree on the rating), the resulting clas-
sification will likely be correct.

EXAMPLE 4 

CART and Ensemble Learning

Laurie Kim is a portfolio manager at Hilux LLC, a high-­yield bond investment
firm. The economy has been in recession for several months, and high-­yield bond
prices have declined precipitously as credit spreads have widened in response
to the weak macroeconomic environment. Kim, however, believes this is a good
time to buy as she expects to profit as credit spreads narrow and high-­yield bond
prices rise in anticipation of economic recovery.
Based on her analysis, Kim believes that corporate high-­yield bonds in the
credit quality range of B/B2 to CCC/Caa2 are the most attractive. However, she
must carefully select which bonds to buy and which bonds to avoid because of
the elevated default risk caused by the currently weak economy.
To help with her bond selection, Kim turns to Hilux’s data analytics team.
Kim has supplied them with historical data consisting of 19 fundamental and 5
technical factors for several thousand high-­yield bond issuers and issues labeled
to indicate default or no default. Kim requests that the team develop an ML-­
based model that will make accurate classifications in two categories: default
and no default. Exploratory data analysis suggests considerable non-­linearities
among the feature set.
1 State the type of problem being addressed by Kim.
2 Describe the dimensionality of the model that Kim requests her analytics
team to develop.
3 Evaluate whether a CART model is appropriate for addressing her
problem.
4 Describe how a CART model operates at each node of the tree.
5 Describe how the team might avoid overfitting and improve the predictive
power of a CART model.
6 Describe how ensemble learning might be used by the team to develop
even better predictions for Kim’s selection of corporate high-­yield bonds.

Solution to 1:
Kim is addressing a classification problem because she must determine whether
bonds that she is considering purchasing in the credit quality range of B/B2 to
CCC/Caa2 will default or not default.
Solution to 2:
With 19 fundamental and 5 technical factors (i.e., the features) the dimension-
ality of the model is 24.
 © CFA Institute. For candidate use only. Not for distribution.
Unsupervised Machine Learning Algorithms 489

Solution to 3:
The CART model is an available algorithm for addressing classification problems.
Its ability to handle complex, non-­linear relationships makes it a good choice
to address the modelling problem at hand. An important advantage of CART
is that its results are relatively straightforward to visualize and interpret, which
should help Kim explain her recommendations based on the model to Hilux’s
investment committee and the firm’s clients.
Solution to 4:
At each node in the decision tree, the algorithm will choose the feature and the
cutoff value for the selected feature that generates the widest separation of the
labeled data to minimize classification error.
Solution to 5:
The team can avoid overfitting and improve the predictive power of the CART
model by adding regularization parameters. For example, the team could spec-
ify the maximum depth of the tree, the minimum population at a node, or the
maximum number of decision nodes. The iterative process of building nodes will
be stopped once the regularization criterion has been reached. Alternatively, a
pruning technique can be used afterward to remove parts of the CART model
that provide little power to correctly classify instances into default or no default
categories.
Solution to 6:
The analytics team might use ensemble learning to combine the predictions from
a collection of models, where the average result of many predictions leads to a
reduction in noise and thus more accurate predictions. Ensemble learning can
be achieved by an aggregation of either heterogeneous learners—different types
of algorithms combined with a voting classifier—or homogenous learners—a
combination of the same algorithm but using different training data based on
the bootstrap aggregating (i.e., bagging) technique. The team may also consider
developing a random forest classifier (i.e., a collection of many decision trees)
trained via a bagging method.

UNSUPERVISED MACHINE LEARNING ALGORITHMS

As you may recall, unsupervised learning is machine learning that does not use labeled
6
data (i.e., no target variable); thus, the algorithms are tasked with finding patterns
within the data themselves. The two main types of unsupervised ML algorithms shown
in Exhibits 2 and 3 are dimension reduction, using principal components analysis,
and clustering, which includes k-­means and hierarchical clustering. These will now
be described in turn.

6.1  Principal Components Analysis

Dimension reduction is an important type of unsupervised learning that is used widely
in practice. When many features are in a data set, representing the data visually or
fitting models to the data may become extremely complex and “noisy” in the sense of
reflecting random influences specific to a data set. In such cases, dimension reduction
may be necessary. Dimension reduction aims to represent a data set with many, typically
correlated, features by a smaller set of features that still do well in describing the data.
 © CFA Institute. For candidate use only. Not for distribution.
490 Reading 7 ■ Machine Learning

A long-­established statistical method for dimension reduction is principal compo-

nents analysis (PCA). PCA is used to summarize or reduce highly correlated features
of data into a few main, uncorrelated composite variables. A composite variable is a
variable that combines two or more variables that are statistically strongly related to
each other. Informally, PCA involves transforming the covariance matrix of the features
and involves two key concepts: eigenvectors and eigenvalues. The eigenvectors define
new, mutually uncorrelated composite variables that are linear combinations of the
original features. As a vector, an eigenvector also represents a direction. Associated
with each eigenvector is an eigenvalue. An eigenvalue gives the proportion of total
variance in the initial data that is explained by each eigenvector. The PCA algorithm
orders the eigenvectors from highest to lowest according to their eigenvalues—that
is, in terms of their usefulness in explaining the total variance in the initial data. PCA
selects as the first principal component the eigenvector that explains the largest pro-
portion of variation in the data set (the eigenvector with the largest eigenvalue). The
second principal component explains the next largest proportion of variation remaining
after the first principal component; it continues for the third, fourth, and subsequent
principal components. As the principal components are linear combinations of the
initial feature set, only a few principal components are typically required to explain
most of the total variance in the initial feature covariance matrix.
Exhibit 13 shows a hypothetical data set with three features, so it is plotted in three
dimensions along the x, y, and z axes. Each data point has a measurement (x, y, z),
and the data should be standardized so that the mean of each series (x’s, y’s, and z’s)
is 0 and the standard deviation is 1. Assume PCA has been applied, revealing the first
two principal components, PC1 and PC2. With respect to PC1, a perpendicular line
dropped from each data point to PC1 shows the vertical distance between the data
point and PC1, representing projection error. Moreover, the distance between each
data point in the direction that is parallel to PC1 represents the spread or variation
of the data along PC1. The PCA algorithm operates in such a way that it finds PC1
by selecting the line for which the sum of the projection errors for all data points is
minimized and for which the sum of the spread between all the data is maximized.
As a consequence of these selection criteria, PC1 is the unique vector that accounts
for the largest proportion of the variance in the initial data. The next largest portion
of the remaining variance is best explained by PC2, which is at right angles to PC1
and thus is uncorrelated with PC1. The data points can now be represented by the
first two principal components. This example demonstrates the effectiveness of the
PCA algorithm in summarizing the variability of the data and the resulting dimension
reduction.
 © CFA Institute. For candidate use only. Not for distribution.
Unsupervised Machine Learning Algorithms 491

Exhibit 13  First and Second Principal Components of a Hypothetical

3-­Dimensional Data Set
Y
PC 2: Spread Projection Error
Eigenvector that
explains the next PC 1
largest % of total
variance remaining.
(after PC 1)

90°
PC 1:
Eigenvector that
X explains the largest
% of total variance
in the data.
PC 2
Z

It is important to know how many principal components to retain because there

is a trade-­off between a lower-­dimensional, more manageable view of a complex data
set when a few are selected and some loss of information. Scree plots, which show
the proportion of total variance in the data explained by each principal component,
can be helpful in this regard (see the accompanying sidebar). In practice, the smallest
number of principal components that should be retained is that which the scree plot
shows as explaining 85% to 95% of total variance in the initial data set.

Scree Plots for the Principal Components of Returns to the

Hypothetical DLC 500 and VLC 30 Equity Indexes
In this illustration, researchers use scree plots and decide that three principal components
are sufficient for explaining the returns to the hypothetical DLC 500 and VLC 30 equity
indexes over the last 10-­year period. The DLC 500 can be thought of as a diversified
index of large-­cap companies covering all economic sectors, while the VLC 30 is a more
concentrated index of the 30 largest publicly traded companies. The data set consists of
index prices and more than 2,000 fundamental and technical features. Multi-­collinearity
among the features is a typical problem because that many features or combinations
of features would tend to have overlaps. To mitigate the problem, PCA can be used to
capture the information and variance in the data. The following scree plots show that
of the 20 principal components generated, the first 3 together explain about 90% and
86% of the variance in the value of the DLC 500 and VLC 30 indexes, respectively. The
scree plots indicate that for each of these indexes, the incremental contribution to
explaining the variance structure of the data is quite small after about the fifth principal
component. Therefore, these less useful principal components can be ignored without
much loss of information.
 © CFA Institute. For candidate use only. Not for distribution.
492 Reading 7 ■ Machine Learning

Scree Plots of Percent of Total Variance Explained by Each Principal

Component for Hypothetical DLC 500 and VLC 30 Equity Indexes
A. Diversified Large Cap 500 Index (DLC 500)
Percent of Variance Explained

0.6

0.4 About 90% of total variance

in DLC 500 Index returns is
explained by PC1 (43%) + PC2
(26%) + PC3 (21%)
0.2

0
0 5 10 15 20
Number of Principal Components

B. Very Large Cap 30 Index (VLC 30)

Percent of Variance Explained

0.6

0.4 About 86% of total variance

in VLC 30 Index returns is
explained by PC1 (55%) + PC2
(20%) + PC3 (11%)
0.2

0
0 5 10 15 20
Number of Principal Components

The main drawback of PCA is that since the principal components are combina-
tions of the data set’s initial features, they typically cannot be easily labeled or directly
interpreted by the analyst. Compared to modelling data with variables that represent
well-­defined concepts, the end user of PCA may perceive PCA as something of a
“black box.”
Reducing the number of features to the most relevant predictors is very useful,
even when working with data sets having as few as ten or so features. Notably, dimen-
sion reduction facilitates visually representing the data in two or three dimensions.
It is typically performed as part of exploratory data analysis, before training another
supervised or unsupervised learning model. Machine learning models are quicker
to train, tend to reduce overfitting (by avoiding the curse of dimensionality), and are
easier to interpret if provided with lower dimensional data sets.

6.2  Clustering
Clustering is another type of unsupervised machine learning that is used to organize
data points into similar groups called clusters. A cluster contains a subset of obser-
vations from the data set such that all the observations within the same cluster are
deemed “similar.” The aim is to find a good clustering of the data—meaning that the
observations inside each cluster are similar or close to each other (a property known
 © CFA Institute. For candidate use only. Not for distribution.
Unsupervised Machine Learning Algorithms 493

as cohesion) and the observations in two different clusters are as far away from one
another or are as dissimilar as possible (a property known as separation). Exhibit 14
depicts this intra-­cluster cohesion and inter-­cluster separation.

Exhibit 14  Evaluating Clustering—Intra-­Cluster Cohesion and Inter-­Cluster

Separation
Bad Clustering Good Clustering (Maybe) Better Clustering

Clustering algorithms are particularly useful in the many investment problems

and applications in which the concept of similarity is important. Applied to grouping
companies, for example, clustering may uncover important similarities and differences
among companies that are not captured by standard classifications of companies by
industry and sector. In portfolio management, clustering methods have been used for
improving portfolio diversification.
In practice, expert human judgment has a role in using clustering algorithms. In
the first place, one must establish what it means to be “similar.” Each company can be
considered an observation with multiple features, including such financial statement
items as total revenue and profit to shareholders, a wide array of financial ratios, or
any other potential model inputs. Based on these features, a measure of similarity
or “distance” between two observations (i.e., companies) can be defined. The smaller
the distance, the more similar the observations; the larger the distance, the more
dissimilar the observations.
A commonly used definition of distance is the Euclidian distance, the straight-­line
distance between two points. Roughly a dozen different distance measures are used
regularly in ML. In practice, the choice of the distance measures depends on the
nature of the data (numerical or not) and the business problem being investigated.
Once the relevant distance measure is defined, similar observations can be grouped
together. We now introduce two of the more popular clustering approaches: k-­means
and hierarchical clustering.

6.2.1  K-­Means Clustering

K-­means is a relatively old algorithm that repeatedly partitions observations into a
fixed number, k, of non-­overlapping clusters. The number of clusters, k, is a model
hyperparameter—a parameter whose value must be set by the researcher before
learning begins. Each cluster is characterized by its centroid (i.e., center), and each
observation is assigned by the algorithm to the cluster with the centroid to which
that observation is closest.
The k-­means algorithm follows an iterative process. It is illustrated in Exhibit 15
for k = 3 and a set of observations on a variable that can be described by two features.
In Exhibit 15, the horizontal and vertical axes represent, respectively, the first and
second characteristics. For example, an investment analyst may want to group a set of
firms into three groups according to two numerical measures of management quality.
The algorithm groups the observations in the following steps:
1 K-­means starts by determining the position of the k (here, 3) initial random
centroids.
 © CFA Institute. For candidate use only. Not for distribution.
494 Reading 7 ■ Machine Learning

2 The algorithm then analyzes the features for each observation. Based on the
distance measure that is utilized, k-­means assigns each observation to its closest
centroid, which defines a cluster.
3 Using the observations within each cluster, k-­means then calculates the new
(k) centroids for each cluster, where the centroid is the average value of their
assigned observations.
4 K-­means then reassigns the observations to the new centroids, redefining the
clusters in terms of included and excluded observations.
5 The process of recalculating the new (k) centroids for each cluster is reiterated.
6 K-­means then reassigns the observations to the revised centroids, again redefin-
ing the clusters in terms of observations that are included and excluded.
The k-­means algorithm will continue to iterate until no observation is reassigned
to a new cluster (i.e., no need to recalculate new centroids). The algorithm has then
converged and reveals the final k clusters with their member observations. The k-­means
algorithm has minimized intra-­cluster distance (thereby maximizing cohesion) and
has maximized inter-­cluster distance (thereby maximizing separation) under the
constraint that k = 3.

Exhibit 15  Example of 3-­Means Algorithm

Step 1: Chooses initial random Step 2: Assigns each obsevation Step 3: Calculates new centroids
centroids: c1, c2, c3 to nearest centroid as the average values of
(defining initial 3 clusters) observations in a cluster

c1 c1 c1
c3 c3 c3

c2 c2 c2

Step 4: Reassigns each Step 5: Reiterates the process of Step 6: Reassigns each observation
observation to the nearest recalculationg new cntroids to the nearest centroid (from Step 5),
centroid (from Step 3) completing second iteration

c1
c1 c3 c1 c3
c3

c2 c2 c2

The k-­means algorithm is fast and works well on very large data sets with hun-
dreds of millions of observations. However, the final assignment of observations to
clusters can depend on the initial location of the centroids. To address this problem,
the algorithm can be run several times using different sets of initial centroids, and
then one can choose the clustering that is most useful given the business purpose.
One limitation of this technique is that the hyperparameter, k, the number of
clusters in which to partition the data, must be decided before k-­means can be run.
So, one needs to have a sense of how many clusters are reasonable for the problem
 © CFA Institute. For candidate use only. Not for distribution.
Unsupervised Machine Learning Algorithms 497

few technical points on dendrograms bear mentioning—although they may not all be
apparent in Exhibit 17. The x-­axis shows the clusters, and the y-­axis indicates some
distance measure. Clusters are represented by a horizontal line, the arch, which con-
nects two vertical lines, called dendrites, where the height of each arch represents
the distance between the two clusters being considered. Shorter dendrites represent
a shorter distance (and greater similarity) between clusters. The horizontal dashed
lines cutting across the dendrites show the number of clusters into which the data
are split at each stage.
The agglomerative algorithm starts at the bottom of the dendrite where each
observation is its own cluster (A to K). Agglomerative clustering then generates the 6
larger clusters (1 to 6). For example, clusters A and B combine to form cluster 1, and
observation G remains its own cluster, now cluster 4. Moving up the dendrogram,
2 larger clusters are formed, where, for example, cluster 7 includes clusters 1 to 3.
Finally, at the top of the dendrogram is the single large cluster (9). The dendrogram
readily shows how this largest cluster is composed of the two main sub-­clusters (7
and 8), each having three smaller sub-­clusters (1 to 3 and 4 to 6, respectively). The
dendrogram also facilitates visualization of divisive clustering by starting at the top
of the largest cluster and then working downward until the bottom is reached where
all 11 single-­observation clusters are shown.

Exhibit 17  Dendrogram of Agglomerative Hierarchical Clustering

.07
Arch

.06 Dendrite
9

.05

.04
7 8

.03

.02
1 3 5 6

.01
2 4

0
A B C D E F G H I J K
Cluster

2 Clusters 6 Clusters 11 Clusters

Clustering has many applications in investment management. For example, portfolio

diversification can be approached as a clustering problem with the aim of optimally
diversifying risks by investing in assets from multiple different clusters. Because the
clusters have maximum inter-­cluster separation, diversifying among them helps ensure
that the portfolio reflects a wide diversity of characteristics with well-­diversified risk.
Alternatively, information that investments are concentrated in a cluster indicates a
high probability of concentrated risk. Finally, it is important to note that while the
results of clustering algorithms are often difficult to evaluate (because the resulting
 © CFA Institute. For candidate use only. Not for distribution.
498 Reading 7 ■ Machine Learning

clusters themselves are not explicitly defined), they are still very useful in practice for
uncovering important underlying structure (namely, similarities among observations)
in complex data sets.

EXAMPLE 5 

Investment Uses of Clustering Algorithms

István Perényi is a portfolio manager of the Europe Diversified Equity Fund (“the
Fund”) within the Diversified Investment Management Company (DIMCO)
fund family. The Fund is benchmarked to the STOXX Europe 600 Index, which
spans 17 countries, 19 industry sectors, and 3 market capitalization groupings
(large-, mid-, and small-­cap).
Examining the Fund’s most recent performance, Perényi is concerned that the
Fund’s holdings, although approximately aligned with the STOXX Europe 600
Index’s country weights, may have unrecognized risk biases and concentrations.
Perényi requests Elsa Lund, DIMCO’s chief risk officer, to investigate the Fund’s
diversification. Lund asks her analysts for ideas on how Perényi’s request can be
addressed and receives three suggestions:
Suggestion 1 Estimate the country, industry, and market-­cap exposures
of each Fund holding, aggregate them, and compare the
aggregate exposures to the benchmark’s exposures. Then,
examine mismatches for evidence of unexpected biases or
concentrations.
Suggestion 2 Identify inherent groupings among fund holdings based
on a broad set of eight numerical (operating and financial)
measures related to the holdings’ characteristics. Then,
examine the groupings for evidence of unexpected biases
or concentrations.
Suggestion 3 Regress the return of the Fund on a set of country equity
market indexes and sector indexes based on the Fund’s
benchmark. Then, examine the regression coefficients for
evidence of unexpected biases or concentrations.
Lund has several questions for analyst Greg Kane about using one or more
clustering machine learning algorithms in relation to addressing Perényi’s request.
Lund asks if any information needs to be specified for ML clustering algo-
rithms no matter which one is used. Kane replies that only the distance measure
that the algorithm will use and the hyperparameter, k, for k-­means clustering
need to be specified.
Lund further asks whether there would be an advantage to using k-­means
clustering as opposed to hierarchical clustering. Kane replies that in his opinion,
hierarchical clustering is the more appropriate algorithm.
1 Which analyst suggestion is most likely to be implemented using machine
learning?
A Suggestion 1.
B Suggestion 2.
C Suggestion 3.
2 Kane’s reply to Lund’s first question is:
A correct.
B not correct, because other hyperparameters must also be specified.
 © CFA Institute. For candidate use only. Not for distribution.
Neural Networks, Deep Learning Nets, and Reinforcement Learning 501

Exhibit 19  A More Complex (4-­5-1) Neural Network with One Hidden Layer

Input Hidden Ouput

Layer Layer Layer

Input 1

Input 2
Ouput
Input 3

Input 4

Now consider any of the nodes to the right of the input layer. These nodes are
sometimes called “neurons” because they process information received. Take the
topmost hidden node. Four links connect to that node from the inputs, so the node
gets four values transmitted by the links. Each node has, conceptually, two functional
parts: a summation operator and an activation function. Once the node receives the
four input values, the summation operator multiplies each value by a weight and sums
the weighted values to form the total net input. The total net input is then passed to
the activation function, which transforms this input into the final output of the node.
Informally, the activation function operates like a light dimmer switch that decreases
or increases the strength of the input. The activation function is characteristically
non-­linear, such as an S-­shaped (sigmoidal) function (with output range of 0 to 1) or
the rectified linear unit function shown in Panel B of Exhibit 18. Non-­linearity implies
that the rate of change of output differs at different levels of input.
This activation function is shown in Exhibit 20, where in the left graph a negative
total net input is transformed via the S-­shaped function into an output close to 0.
This low output implies the node does not “fire,” so there is nothing to pass to the
next node. Conversely, in the right graph a positive total net input is transformed
into an output close to 1, so the node does fire. The output of the activation function
is then transmitted to the next set of nodes if there is a second hidden layer or, as in
this case, to the output layer node as the predicted value. The process of transmis-
sion just described (think of forward pointing arrows in Exhibit 19) is referred to as
forward propagation.
 © CFA Institute. For candidate use only. Not for distribution.
502 Reading 7 ■ Machine Learning

Exhibit 20  Activation Function as “Light Dimmer Switch” at Each Node in a

Neural Network
Activation Function Output Activation Function Output
1 1

Sigmoid Sigmoid
Function Function

0 0
– Summation Operator + – Summation Operator +
Total Net Input Total Net Input
Dimmer Dimmer
Switch Switch

Starting with an initialized set of random network weights, training a neural net-
work in a supervised learning context is an iterative process in which predictions are
compared to actual values of labeled data and evaluated by a specified performance
measure (e.g., mean squared error). Then, network weights are adjusted to reduce
total error of the network. (If the process of adjustment works backward through the
layers of the network, this process is called backward propagation). Learning takes
place through this process of adjustment to network weights with the aim of reducing
total error. Without proliferating notation relating to nodes, the gist of the updating
can be expressed informally as:

New weight = (Old weight) – (Learning rate) × (Partial derivative of the total
error with respect to the old weight),
where “partial derivative” is a “gradient” or “rate of change of the total error with
respect to the change in the old weight,” and learning rate is a parameter that affects
the magnitude of adjustments. When learning is complete, all the network weights
have assigned values.
The structure of a network in which all the features are interconnected with
non-­linear activation functions allows neural networks to uncover and approximate
complex non-­linear relationships among features. Broadly speaking, when more nodes
and more hidden layers are specified, a neural network’s ability to handle complexity
tends to increase (but so does the risk of overfitting).
Asset pricing is a noisy, stochastic process with potentially unstable relationships
that challenge modelling processes, so researchers are asking if machine learning
can improve our understanding of how markets work. Research comparing statisti-
cal and machine learning methods’ abilities to explain and predict equity prices so
far indicates that simple neural networks produce models of equity returns at the
individual stock and portfolio level that are superior to models built using traditional
statistical methods due to their ability to capture dynamic and interacting variables.
This suggests that ML-­based models, such as neural networks, may simply be better
able to cope with the non-­linear relationships inherent in security prices. However,
the trade-­offs in using them are the lack of interpretability and the amount of data
needed to train such models.
 © CFA Institute. For candidate use only. Not for distribution.
Neural Networks, Deep Learning Nets, and Reinforcement Learning 505

Solution to 1:
A deep learning net (DLN) is a neural network (NN) with many hidden layers (at
least 3 but often more than 20). NNs and DLNs have been successfully applied to
a wide variety of complex tasks characterized by non-­linearities and interactions
among features, particularly pattern recognition problems.
Solution to 2:
Mitsui wants to detect patterns of potential style drift in the daily trading of
nearly 100 external asset managers in many markets. This task will involve the
fast processing of huge amounts of complicated data. Monroe is correct that a
DLN is well suited to PEPF’s needs.
Solution to 3:
The input layer, the hidden layers, and the output layer constitute the three
groups of layers of DLNs. The input layer receives the inputs (i.e., features) and
has as many nodes as there are dimensions of the feature set. The hidden layers
consist of nodes, each comprised of a summation operator and an activation
function that are connected by links. These hidden layers are, in effect, where the
model is learned. The final layer, the output layer, produces a set of probabilities
of an observation being in any of the target style categories (each represented
by a node in the output layer). The DLN assigns the category based on the style
category with the highest probability.

EXAMPLE 7 

Summing Up the Major Types of Machine Learning

1 As used in supervised machine learning, regression problems involve:
A binary target variables.
B continuous target variables.
C categorical target variables.
2 Which of the following best describes penalized regression? Penalized
regression:
A is unrelated to multiple linear regression.
B involves a penalty term that is added to the predicted target variable.
C is a category of general linear models used when the number of fea-
tures and overfitting are concerns.
3 CART is best described as:
A an unsupervised ML algorithm.
B a clustering algorithm based on decision trees.
C a supervised ML algorithm that accounts for non-­linear relationships
among the features.
4 A neural network is best described as a technique for machine learning
that is:
A exactly modeled on the human nervous system.
B based on layers of nodes connected by links when the relationships
among the features are usually non-­linear.
 © CFA Institute. For candidate use only. Not for distribution.
506 Reading 7 ■ Machine Learning

C based on a tree structure of nodes when the relationships among the

features are linear.
5 Hierarchical clustering is best described as a technique in which:
A the grouping of observations is unsupervised.
B features are grouped into a pre-­specified number, k, of clusters.
C observations are classified according to predetermined labels.
6 Dimension reduction techniques are best described as a means to reduce a
set of features:
A to a manageable size without regard for the variation in the data.
B to a manageable size while increasing the variation in the data.
C to a manageable size while retaining as much of the variation in the
data as possible.

Solution to 1:
B is correct. A and C are incorrect because when the target variable is binary
or categorical, the problem is a classification problem rather than a regression
problem.
Solution to 2:
C is correct. A is incorrect because penalized regression is related to multiple
linear regression. B is incorrect because penalized regression involves adding a
penalty term to the sum of the squared regression residuals.
Solution to 3:
C is correct. A is incorrect because CART is a supervised ML algorithm. B
is incorrect because CART is a classification and regression algorithm, not a
clustering algorithm.
Solution to 4:
B is correct. A is incorrect because neural networks are not exactly modeled on
the human nervous system. C is incorrect because neural networks are not based
on a tree structure of nodes when the relationships among the features are linear.
Solution to 5:
A is correct. B is incorrect because it refers to k-­means clustering. C is incorrect
because it refers to classification, which involves supervised learning.
Solution to 6:
C is correct because dimension reduction techniques, like PCA, are aimed at
reducing the feature set to a manageable size while retaining as much of the
variation in the data as possible.
 © CFA Institute. For candidate use only. Not for distribution.
References 509

■■ Neural networks with many hidden layers (at least 3 but often more than 20)
are known as deep learning nets (DLNs) and are the backbone of the artificial
intelligence revolution.
■■ The RL algorithm involves an agent that should perform actions that will
maximize its rewards over time, taking into consideration the constraints of its
environment.

REFERENCES
Bacham, Dinesh, and Janet Zhao. 2017. “Machine Learning:
Challenges, Lessons, and Opportunities in Credit Risk
Modeling.” Moody’s Analytics Risk Perspectives 9:28–35.
 © CFA Institute. For candidate use only. Not for distribution.
510 Reading 7 ■ Machine Learning

PRACTICE PROBLEMS

The following information relates to Questions

1–10
Alef Associates manages a long-­only fund specializing in global smallcap equities. Since
its founding a decade ago, Alef maintains a portfolio of 100 stocks (out of an eligible
universe of about 10,000 stocks). Some of these holdings are the result of screening
the universe for attractive stocks based on several ratios that use readily available
market and accounting data; others are the result of investment ideas generated by
Alef ’s professional staff of five securities analysts and two portfolio managers.
Although Alef ’s investment performance has been good, its Chief Investment
Officer, Paul Moresanu, is contemplating a change in the investment process aimed
at achieving even better returns. After attending multiple workshops and being
approached by data vendors, Moresanu feels that data science should play a role in
the way Alef selects its investments. He has also noticed that much of Alef ’s past
outperformance is due to stocks that became takeover targets. After some research
and reflection, Moresanu writes the following email to the Alef ’s CEO.

Subject: Investment Process Reorganization

I have been thinking about modernizing the way we select stock investments.
Given that our past success has put Alef Associates in an excellent financial
position, now seems to be a good time to invest in our future. What I propose
is that we continue managing a portfolio of 100 global small-­cap stocks but
restructure our process to benefit from machine learning (ML). Importantly, the
new process will still allow a role for human insight, for example, in providing
domain knowledge. In addition, I think we should make a special effort to iden-
tify companies that are likely to be acquired. Specifically, I suggest following the
four steps which would be repeated every quarter.
Step 1 We apply ML techniques to a model including fundamental and
technical variables (features) to predict next quarter’s return
for each of the 100 stocks currently in our portfolio. Then, the
20 stocks with the lowest estimated return are identified for
replacement.
Step 2 We utilize ML techniques to divide our investable universe of
about 10,000 stocks into 20 different groups, based on a wide
variety of the most relevant financial and non-­financial character-
istics. The idea is to prevent unintended portfolio concentration
by selecting stocks from each of these distinct groups.
Step 3 For each of the 20 different groups, we use labeled data to train
a model that will predict the five stocks (in any given group) that
are most likely to become acquisition targets in the next one year.

© 2019 CFA Institute. All rights reserved.

 © CFA Institute. For candidate use only. Not for distribution.
Practice Problems 513

B The activation function in a node operates like a light dimmer switch since
it decreases or increases the strength of the total net input.
C The summation operator receives input values, multiplies each by a weight,
sums up the weighted values into the total net input, and passes it to the
activation function.
 © CFA Institute. For candidate use only. Not for distribution.
514 Reading 7 ■ Machine Learning

SOLUTIONS
1 A is correct. The target variable (quarterly return) is continuous, hence this calls
for a supervised machine learning based regression model.
B is incorrect, since classification uses categorical or ordinal target variables,
while in Step 1 the target variable (quarterly return) is continuous.
C is incorrect, since clustering involves unsupervised machine learning so does
not have a target variable.
2 B is correct. It is least appropriate because with LASSO, when λ = 0 the penalty
(i.e., regularization) term reduces to zero, so there is no regularization and the
regression is equivalent to an ordinary least squares (OLS) regression.
A is incorrect. With Classification and Regression Trees (CART), one way that
regularization can be implemented is via pruning which will reduce the size of
the regression tree—sections that provide little explanatory power are pruned
(i.e., removed).
C is incorrect. With LASSO, when λ is between 0.5 and 1 the relatively large
penalty (i.e., regularization) term requires that a feature makes a sufficient con-
tribution to model fit to offset the penalty from including it in the model.
3 A is correct. K-­Means clustering is an unsupervised machine learning algo-
rithm which repeatedly partitions observations into a fixed number, k, of non-­
overlapping clusters (i.e., groups).
B is incorrect. Principal Components Analysis is a long-­established statistical
method for dimension reduction, not clustering. PCA aims to summarize or
reduce highly correlated features of data into a few main, uncorrelated compos-
ite variables.
C is incorrect. CART is a supervised machine learning technique that is most
commonly applied to binary classification or regression.
4 C is correct. Here, 20 is a hyperparameter (in the K-­Means algorithm), which is
a parameter whose value must be set by the researcher before learning begins.
A is incorrect, because it is not a hyperparameter. It is just the size (number of
stocks) of Alef ’s portfolio.
B is incorrect, because it is not a hyperparameter. It is just the size (number of
stocks) of Alef ’s eligible universe.
5 B is correct. To predict which stocks are likely to become acquisition targets,
the ML model would need to be trained on categorical labelled data having the
following two categories: “0” for “not acquisition target”, and “1” for “acquisition
target”.
A is incorrect, because the target variable is categorical, not continuous.
C is incorrect, because the target variable is categorical, not ordinal (i.e., 1st,
2nd, 3rd, etc.).
6 C is correct. The advantages of using CART over KNN to classify companies
into two categories (“not acquisition target” and “acquisition target”), include
all of the following: For CART there are no requirements to specify an initial
hyperparameter (like K) or a similarity (or distance) measure as with KNN, and
CART provides a visual explanation for the prediction (i.e., the feature variables
and their cut-­off values at each node).
A is incorrect, because CART provides all of the advantages indicated in
Statements I, II and III.
 © CFA Institute. For candidate use only. Not for distribution.

READING

8
Big Data Projects
by Sreekanth Mallikarjun, PhD, and Ahmed Abbasi, PhD
Sreekanth Mallikarjun, PhD, is at Reorg (USA) and the University of Virginia, McIntire
School of Commerce (USA). Ahmed Abbasi, PhD, is at the University of Virginia, McIntire
School of Commerce (USA).

LEARNING OUTCOMES
Mastery The candidate should be able to:

a. state and explain steps in a data analysis project;

b. describe objectives, steps, and examples of preparing and
wrangling data;
c. describe objectives, methods, and examples of data exploration;
d. describe objectives, steps, and techniques in model training;
e. describe preparing, wrangling, and exploring text-­based data for
financial forecasting;
f. describe methods for extracting, selecting and engineering
features from textual data;
g. evaluate the fit of a machine learning algorithm.

INTRODUCTION
Big data (also referred to as alternative data) encompasses data generated by financial
1
markets (e.g., stock and bond prices), businesses (e.g., company financials, production
volumes), governments (e.g., economic and trade data), individuals (e.g., credit card
purchases, social media posts), sensors (e.g., satellite imagery, traffic patterns), and
the Internet of Things, or IoT, (i.e., the network of interrelated digital devices that can
transfer data among themselves without human interaction). A veritable explosion in
big data has occurred over the past decade or so, especially in unstructured data gen-
erated from social media (e.g., posts, tweets, blogs), email and text communications,
web traffic, online news sites, electronic images, and other electronic information
sources. The prospects are for exponential growth in big data to continue.

© 2019 CFA Institute. All rights reserved.

 © CFA Institute. For candidate use only. Not for distribution.
518 Reading 8 ■ Big Data Projects

Investment managers are increasingly using big data in their investment processes
as they strive to discover signals embedded in such data that can provide them with an
information edge. They seek to augment structured data with a plethora of unstruc-
tured data to develop improved forecasts of trends in asset prices, detect anomalies,
etc. A typical example involves a fund manager using financial text data from 10-­K
reports for forecasting stock sentiment (i.e., positive or negative), which can then be
used as an input to a more comprehensive forecasting model that includes corporate
financial data.
Unlike structured data (numbers and values) that can be readily organized into
data tables to be read and analyzed by computers, unstructured data typically require
specific methods of preparation and refinement before being usable by machines (i.e.,
computers) and useful to investment professionals. Given the volume, variety, and
velocity of available big data, it is important for portfolio managers and investment
analysts to have a basic understanding of how unstructured data can be transformed
into structured data suitable as inputs to machine learning (ML) methods (in fact, for
any type of modeling methods) that can potentially improve their financial forecasts.
This reading describes the steps in using big data, both structured and unstruc-
tured, in financial forecasting. The concepts and methods are then demonstrated in a
case study of an actual big data project. The project uses text-­based data derived from
financial documents to train an ML model to classify text into positive or negative
sentiment classes for the respective stocks and then to predict sentiment.
Section 2 of the reading covers a description of the key characteristics of big data.
Section 3 provides an overview of the steps in executing a financial forecasting project
using big data. We then describe in Sections 4–6 key aspects of data preparation and
wrangling, data exploration, and model training using structured data and unstructured
(textual) data. In Section 7, we bring these pieces together by covering the execution
of an actual big data project. A summary in Section 8 concludes the reading.

2 BIG DATA IN INVESTMENT MANAGEMENT

Big data differs from traditional data sources based on the presence of a set of char-
acteristics commonly referred to as the 3Vs: volume, variety, and velocity.
Volume refers to the quantity of data. The US Library of Congress, which is tasked
with archiving both digital and physical information artifacts in the United States, has
collected hundreds of terabytes of data (one terabyte equals 1,024 gigabytes, which
are equal to 1,048,576 megabytes). Several years ago, one of the authors managed an
archival project for the Library of Congress in which many terabytes of online content
were collected—a copious amount of data at the time. However, in most US industry
sectors today, the average company collects more data than the Library of Congress!
In big data conversations, terabytes have been replaced with petabytes and exabytes
(one exabyte equals 1,024 petabytes, which are equal to 1,048,576 terabytes). The
classic grains of sand analogy puts these volumes into perspective: If a megabyte is a
tablespoon of sand, then a petabyte is a 1.6-­kilometer-­long beach and an exabyte is a
beach extending about 1,600 kilometers.
Variety pertains to the array of available data sources. Organizations are now deal-
ing with structured, semi-­structured, and unstructured data from within and outside
the enterprise. Variety includes traditional transactional data; user-­generated text,
images, and videos; social media; sensor-­based data; web and mobile clickstreams; and
spatial-­temporal data. Effectively leveraging the variety of available data presents both
opportunities and challenges, including such legal and ethical issues as data privacy.
 © CFA Institute. For candidate use only. Not for distribution.
Steps in Executing a Data Analysis Project: Financial Forecasting with Big Data 521

3 Text preparation and wrangling. This step involves critical cleansing and
preprocessing tasks necessary to convert streams of unstructured data into a
format that is usable by traditional modeling methods designed for structured
inputs.
4 Text exploration. This step encompasses text visualization through techniques,
such as word clouds, and text feature selection and engineering.
The resulting output (e.g., sentiment prediction scores) can either be combined
with other structured variables or used directly for forecasting and/or analysis.
Next, we describe two key steps from the ML Model Building Steps depicted in
Exhibit  1 that typically differ for structured data versus textual big data: data/text
preparation and wrangling and data/text exploration. We then discuss model training.
Finally, we focus on applying these steps to a case study related to classifying and
predicting stock sentiment from financial texts.

Exhibit 1  Model Building for Financial Forecasting Using Big Data:

Structured (Traditional) vs. Unstructured (Text)
ML Model Building Steps

Structured Data
Structured Data Input Collection

Data Prep & Structured

Conceptualization
Wrangling Data
Internal External Classifier
Sources Sources Model Data Output
Training Exploration
Financial
Text Classifier Output Forecasting
Unstructured and
(Structured)
Input Analysis
Data
Curation Text
News Articles Social Media Classifier
Text Problem Text Prep & Output
Formulation Wrangling

Other Open Data Model Text

Documents Training Exploration
Unstructured (Text) Data
Text ML Model Building Steps

EXAMPLE 1 

Steps in ML Model Building

LendALot Corporation is a B2C (business-­to-­consumer) lender that has tradi-
tionally outsourced potential customers’ creditworthiness scoring to a third-­party
firm. Given the recent advances in machine learning (ML)-based “fintech” that
goes beyond traditional “repayment history” and “ability to repay” assessments
derived from structured data, LendALot would like to develop in-­house, ML-­
based credit scoring capabilities to enhance borrower risk assessment and
differentiate itself in the B2C lending market. LendALot would like to follow a
phased approach beginning with traditional (structured) data sources and then
eventually incorporating textual (unstructured) big data sources. Paul Wang has
 © CFA Institute. For candidate use only. Not for distribution.
522 Reading 8 ■ Big Data Projects

been asked to lead a new analytics team at LendALot tasked with developing the
ML-­based creditworthiness scoring model. In the context of machine learning
using structured data sources, address the following questions.
1 State and explain one decision Wang will need to make related to:
A conceptualizing the modeling task.
B data collection.
C data preparation and wrangling.
D data exploration.
E model training.
In a later phase of the project, LendALot attempts to improve its credit
scoring processes by incorporating textual data in credit scoring. Wang tells his
team, “Enhance the creditworthiness scoring model by incorporating insights
from text provided by the prospective borrowers in the loan application free
response fields.”
2 Identify the process step that Wang’s statement addresses.
3 State two potential needs of the LendAlot team in relation to text
curation.
4 State two potential needs of the LendAlot team in relation to text prepara-
tion and wrangling.

Solution to 1:

A In the conceptualization step, Wang will need to decide how the output
of the ML model will be specified (e.g., a binary classification of credit-
worthiness), how the model will be used and by whom, and how it will be
embedded in LendALot’s business processes.
B In the data collection phase, Wang must decide on what data—internal,
external, or both—to use for credit scoring.
C In the data preparation and wrangling step, Wang will need to decide on
data cleansing and preprocessing needs. Cleansing may entail resolving
missing values, extreme values, etc. Preprocessing may involve extracting,
aggregating, filtering, and selecting relevant data columns.
D In the data exploration phase, Wang will need to decide which exploratory
data analysis methods are appropriate, which features to use in building a
credit scoring model, and which features may need to be engineered.
E In the model training step, Wang must decide which ML algorithm(s)
to use. Assuming labeled training data are available, the choice will be
among supervised learning algorithms. Decisions will need to be made on
how model fit is measured and how the model is validated and tuned.

Solution to 2:
Wang’s statement relates to the initial step of text problem formulation.
Solution to 3:
Related to text curation, the team will be using internal data (from loan appli-
cations). They will need to ensure that the text comment fields on the loan
applications have been correctly implemented and enabled. If these fields are
not required, they need to ensure there is a sufficient response rate to analyze.
 © CFA Institute. For candidate use only. Not for distribution.
Data Preparation and Wrangling 525

Exhibit 3  (Continued)

Other
1 ID Name Gender Date of Birth Salary Income State Credit Card
6 5 Ms. XYZ F 15/1/1975 $60,500 $0 Y
7 6 Mr. GHI M 9/10/1942 NA $55,000 TX N
8 7 Mr. TUV M 2/27/1956 $300,000 $50,000 CT Y
9 8 Ms. DEF F 4/4/1980 $55,000 $0 British N
Columbia

The possible errors in a raw dataset include the following:

1 Incompleteness error is where the data are not present, resulting in missing data.
This can be corrected by investigating alternate data sources. Missing values
and NAs (not applicable or not available values) must be either omitted or
replaced with “NA” for deletion or substitution with imputed values during the
data exploration stage. The most common imputations are mean, median, or
mode of the variable or simply assuming zero. In Exhibit 3, rows 4 (ID 3), 5 (ID
4), 6 (ID 5), and 7 (ID 6) are incomplete due to missing values in either Gender,
Salary, Other Income, Name (Salutation), and State columns.
2 Invalidity error is where the data are outside of a meaningful range, resulting
in invalid data. This can be corrected by verifying other administrative data
records. In Exhibit 3, row 5 likely contains invalid data as the date of birth is out
of the range of the expected human life span.
3 Inaccuracy error is where the data are not a measure of true value. This can be
rectified with the help of business records and administrators. In Exhibit 3, row
5 is inaccurate (it shows “Don’t Know”); in reality, every person either has a
credit card or does not.
4 Inconsistency error is where the data conflict with the corresponding data points
or reality. This contradiction should be eliminated by clarifying with another
source. In Exhibit 3, row 3 (ID 2) is likely to be inconsistent as the Name col-
umn contains a female title and the Gender column contains male.
5 Non-­uniformity error is where the data are not present in an identical format.
This can be resolved by converting the data points into a preferable standard
format. In Exhibit 3, the data under the Date of Birth column is present in
various formats. The data under the Salary column may also be non-­uniform as
the monetary units are ambiguous; the dollar symbol can represent US dollar,
Canadian dollar, or others.
6 Duplication error is where duplicate observations are present. This can be
corrected by removing the duplicate entries. In Exhibit 3, row 6 is a duplicate as
the data under Name and Date of Birth columns are identical to the ones in row
3, referring to the same customer.
Exhibit 4 shows the dataset after completion of the cleansing process.
 © CFA Institute. For candidate use only. Not for distribution.
526 Reading 8 ■ Big Data Projects

Exhibit 4.  Data After Cleansing

Other
1 ID Name Gender Date of Birth Salary Income State Credit Card

2 1 Mr. ABC M 12/5/1970 USD 50200 USD 5000 VA Y

3 2 Ms. XYZ F 1/15/1975 USD 60500 USD 0 NY Y
4 3 Mr. EFG M 1/13/1979 USD 65000 USD 1000 CA N
5 6 Mr. GHI M 9/10/1942 USD 0 USD 55000 TX N
6 7 Mr. TUV M 2/27/1956 USD 300000 USD 50000 CT Y
7 8 Ms. DEF F 4/4/1980 CAD 55000 CAD 0 British N
Columbia

Data cleansing can be expensive and cumbersome because it involves the use of
automated, rule-­based, and pattern recognition tools coupled with manual human
inspection to sequentially check for the aforementioned types of errors row by row
and column by column. The process involves a detailed data analysis as an initial
step in identifying various errors that are present in the data. In addition to a manual
inspection and verification of the data, analysis software, such as SPSS, can be used
to understand metadata (data that describes and gives information about other data)
about the data properties to use as a starting point to investigate any errors in the data.
The business value of the project determines the necessary quality of data cleansing
and subsequently the amount of resources used in the cleansing process. In case the
errors cannot be resolved due to lack of available resources, the data points with errors
can simply be omitted depending on the size of the dataset. For instance, if a dataset is
large with more than 10,000 rows, removing a few rows (approximately 100) may not
have a significant impact on the project. If a dataset is small with less than 1,000 rows,
every row might be important and deleting many rows thus harmful to the project.

Data Wrangling (Preprocessing)

To make structured data ready for analyses, the data should be preprocessed. Data
preprocessing primarily includes transformations and scaling of the data. These
processes are exercised on the cleansed dataset. The following transformations are
common in practice:
1 Extraction: A new variable can be extracted from the current variable for ease
of analyzing and using for training the ML model. In Exhibit 4, the Date of
Birth column consists of dates that are not directly suitable for analyses. Thus,
an additional variable called “Age” can be extracted by calculating the number
of years between the present day and date of birth.
2 Aggregation: Two or more variables can be aggregated into one variable to
consolidate similar variables. In Exhibit 4, the two forms of income, Salary and
Other Income, can be summed into a single variable called Total Income.
3 Filtration: The data rows that are not needed for the project must be identified
and filtered. In Exhibit 4, row 7 (ID 8) has a non-­US state; however, this dataset
is for the US-­based bank customers where it is required to have a US address.
 © CFA Institute. For candidate use only. Not for distribution.
Data Preparation and Wrangling 529

Statement 2 “Let’s create a new feature, “Loan Amount as a Percent of

Income,” to use as an additional feature.”

1 The data shown for Ms. Beta contain what is best described as an:
A invalidity error.
B inaccuracy error.
C incompleteness error.
2 The data shown for Mr. Gamma contain what is best described as an:
A invalidity error.
B duplication error.
C incompleteness error.
3 The data shown for Ms. Delta contain what is best described as an:
A invalidity error.
B inaccuracy error.
C duplication error.
4 The data shown for Mr. Zeta contain what is best described as an:
A invalidity error.
B inaccuracy error.
C duplication error.
5 The process mentioned in Wang’s first statement is best described as:
A feature selection.
B feature extraction.
C feature engineering
6 Wang’s second statement is best described as:
A feature selection.
B feature extraction.
C feature engineering.

Solution to 1:
A is correct. This is an invalidity error because the data are outside of a mean-
ingful range. Income cannot be negative.
Solution to 2:
C is correct. This is an incompleteness error as the loan type is missing.
Solution to 3:
B is correct. This is an inaccuracy error because LendALot must know how
much they have lent to that particular borrower (who eventually repaid the loan
as indicated by the loan outcome of no default).
Solution to 4:
C is correct. Row 8 duplicates row 7: This is a duplication error.
Solution to 5:
A is correct. The process mentioned involves selecting the features to use. The
proposal makes sense; with “ID,” “Name” is not needed to identify an observation.
 © CFA Institute. For candidate use only. Not for distribution.
530 Reading 8 ■ Big Data Projects

Solution to 6:
B is correct. The proposed feature is a ratio of two existing features. Feature
extraction is the process of creating (i.e., extracting) new variables from existing
ones in the data.

4.2  Unstructured (Text) Data

Unstructured data are not organized into any systematic format that can be processed
by computers directly. They are available in formats meant for human usage rather
than computer processing. Unstructured data constitute approximately 80% of the
total data available today. They can be in the form of text, images, videos, and audio
files. Unlike in structured data, preparing and wrangling unstructured data are both
more challenging. For analysis and use to train the ML model, the unstructured data
must be transformed into structured data. In this section, text data will be used to
demonstrate unstructured data preparation and wrangling. The cleansing and prepro-
cessing of text data is called text processing. Text processing is essentially cleansing
and transforming the unstructured text data into a structured format. Text processing
can be divided into two tasks: cleansing and preprocessing. The following content is
related to text data in the English language.

Text Preparation (Cleansing)

Raw text data are a sequence of characters and contain other non-­useful elements,
including html tags, punctuations, and white spaces (including tabs, line breaks,
and new lines). It is important to clean the text data before preprocessing. Exhibit 6
shows a sample text from the home page for the hypothetical company Robots Are Us
website. The text appears to be clean visually and is designed for human readability.

Exhibit 6  Sample Text from Robots Are Us Home Page

Robots Are Us

Every home and

business should
have a robot.

However, the source text that can be downloaded is not as clean. The raw text
contains html tags and formatting elements along with the actual text. Exhibit 7 shows
the raw text from the source.

Exhibit 7  Raw Text from the Source

<h1 class=”text-left mb-3”>Robots Are Us</h1>
<h2> Every home and business shoudl have a robot </h2>

The initial step in text processing is cleansing, which involves basic operations to
clean the text by removing unnecessary elements from the raw text. Text operations
often use regular expressions. A regular expression (regex) is a series that contains
characters in a particular order. Regex is used to search for patterns of interest in a
given text. For example, a regex “<.*?>” can be used to find all the html tags that are
 © CFA Institute. For candidate use only. Not for distribution.
Data Preparation and Wrangling 533

Exhibit 9  Tokenization of Four Texts

Cleaned Texts Tokens

Text 1 The man went to the market today The man went to the market today

Text 2 Market values are increasing Market values are increasing

Text 3 Increased marketing is needed Increased marketing is needed

Text 4 There is no market for the product There is no market for the product

Similar to structured data, text data also require normalization. The normalization
process in text processing involves the following:
1 Lowercasing the alphabet removes distinctions among the same words due to
upper and lower cases. This action helps the computers to process the same
words appropriately (e.g., “The” and “the”).
2 Stop words are such commonly used words as “the,” “is,” and “a.” Stop words do
not carry a semantic meaning for the purpose of text analyses and ML training.
However, depending on the end-­use of text processing, for advance text applica-
tions it may be critical to keep the stop words in the text in order to understand
the context of adjacent words. For ML training purposes, stop words typically
are removed to reduce the number of tokens involved in the training set. A
predefined list of stop words is available in programming languages to help with
this task. In some cases, additional stop words can be added to the list based
on the content. For example, the word “exhibit” may occur often in financial
filings, which in general is not a stop word but in the context of the filings can
be treated as a stop word.
3 Stemming is the process of converting inflected forms of a word into its base
word (known as stem). Stemming is a rule-­based approach, and the results need
not necessarily be linguistically sensible. Stems may not be the same as the
morphological root of the word. Porter’s algorithm is the most popular method
for stemming. For example, the stem of the words “analyzed” and “analyzing”
is “analyz.” Similarly, the British English variant “analysing” would become
“analys.” Stemming is available in R and Python. The text mining package in R
provides a stemDocument function that uses this algorithm.
4 Lemmatization is the process of converting inflected forms of a word into
its morphological root (known as lemma). Lemmatization is an algorithmic
approach and depends on the knowledge of the word and language structure.
For example, the lemma of the words “analyzed” and “analyzing” is “analyze.”
Lemmatization is computationally more expensive and advanced.
Stemming or lemmatization will reduce the repetition of words occurring in var-
ious forms and maintain the semantic structure of the text data. Stemming is more
common than lemmatization in the English language since it is simpler to perform.
In text data, data sparseness refers to words that appear very infrequently, resulting
in data consisting of many unique, low frequency tokens. Both techniques decrease
data sparseness by aggregating many sparsely occurring words in relatively less sparse
stems or lemmas, thereby aiding in training less complex ML models.
 © CFA Institute. For candidate use only. Not for distribution.
534 Reading 8 ■ Big Data Projects

After the cleansed text is normalized, a bag-­of-­words is created. Bag-­of-­words

(BOW) representation is a basic procedure used to analyze text. It is essentially a col-
lection of a distinct set of tokens from all the texts in a sample dataset. BOW is simply
a set of words and does not capture the position or sequence of words present in the
text. However, it is memory efficient and easy to handle for text analyses.
Exhibit 10 shows the BOW and transformations occurring in each step of normal-
ization on the cleansed texts from Exhibit 9. Note that the number of words decreases
as the normalizing steps are applied, making the resulting BOW smaller and simpler.

Exhibit 10  Bag-­of-­Words Representation of Four Texts Before and After

Normalization Process
BOW before normalizing
“The” “man” “went” “to” “the” “market”
“today” “Market” “values” “are” “increasing” “Increased”
“marketing” “is” “needed” “There” “no” “for”
“product”

BOW after removing uppercase letters

“the” “man” “went” “to” “market” “today”
“values” “are” “increasing” “increased” “marketing” “is”
“needed” “there” “no” “for” “product”

BOW after removing stop words

“man” “went” “market” “today” “values” “increasing”
“increased” “marketing” “needed” “product”

BOW after stemming

“man” “went” “market” “today” “valu” “increas” “need” “product”

The last step of text preprocessing is using the final BOW after normalizing to build
a document term matrix (DTM). DTM is a matrix that is similar to a data table for
structured data and is widely used for text data. Each row of the matrix belongs to a
document (or text file), and each column represents a token (or term). The number of
rows of DTM is equal to the number of documents (or text files) in a sample dataset.
The number of columns is equal to the number of tokens from the BOW that is built
using all the documents in a sample dataset. The cells can contain the counts of the
number of times a token is present in each document. The matrix cells can be filled
with other values that will be explained in the financial forecasting project section of
this reading; a large dataset is helpful in understanding the concepts. At this point,
the unstructured text data are converted to structured data that can be processed
further and used to train the ML model. Exhibit 11 shows a DTM constructed from
the resultant BOW of the four texts from Exhibit 10.

Exhibit 11  DTM of Four Texts and Using Normalized BOW Filled with Counts
of Occurrence
man went market today valu increas need product

Text 1 1 1 1 1 0 0 0 0
Text 2 0 0 1 0 1 1 0 0
Text 3 0 0 1 0 0 1 1 0
Text 4 0 0 1 0 0 0 0 1
 © CFA Institute. For candidate use only. Not for distribution.
Data Exploration Objectives and Methods 537

Solution to 2:
C is correct. Some punctuations, such as percentage signs, currency symbols,
and question marks, may be useful for ML model training, so when such punc-
tuations are removed annotations should be added.
Solution to 3:
A is correct. Each column of a document term matrix represents a token from
the bag-­of-­words that is built using all the documents in a sample dataset.
Solution to 4:
B is correct. A cell in a document term matrix contains a count of the number
of tokens of the kind indicated in the column heading.
Solution to 5:
C is correct. The other choices are related to text cleansing.
Solution to 6:
A is correct. Stemming, the process of converting inflected word forms into a
base word (or stem), is one technique that can address the problem described.
Solution to 7:
C is correct, by definition. The other choices are not true.

DATA EXPLORATION OBJECTIVES AND METHODS

Data exploration is a crucial part of big data projects. The prepared data are explored
5
to investigate and comprehend data distributions and relationships. The knowledge
that is gained about the data in this stage is used throughout the project. The outcome
and quality of exploration strongly affects ML model training results. Domain knowl-
edge plays a vital role in exploratory analysis as this stage should involve cooperation
between analysts, model designers, and experts in the particular data domain. Data
exploration without domain knowledge can result in ascertaining spurious relationships
among the variables in the data that can mislead the analyses. The data exploration
stage follows the data preparation stage and leads to the model training stage.
Data exploration involves three important tasks: exploratory data analysis, feature
selection, and feature engineering. These three tasks are outlined in Exhibit 13 and are
defined and further explained under the structured and unstructured data subsections.

Exhibit 13  Data Exploration Stage

Data Exploration
Data Data
Collection/ Preparation Exploratory Model
and Feature Feature Results
Curation Data Training
Wrangling Selection Engineering
Analysis

Exploratory data analysis (EDA) is the preliminary step in data exploration.

Exploratory graphs, charts, and other visualizations, such as heat maps and word clouds,
are designed to summarize and observe data. In practice, many exploratory graphs
are made for investigation and can be made swiftly using statistical programming and
generic spreadsheet software tools. Data can also be summarized and examined using
 © CFA Institute. For candidate use only. Not for distribution.
538 Reading 8 ■ Big Data Projects

quantitative methods, such as descriptive statistics and central tendency measures. An

important objective of EDA is to serve as a communication medium among project
stakeholders, including business users, domain experts, and analysts. Relatively quick
and easy exploratory visualizations help stakeholders connect and ensure the prepared
data are sensible. Other objectives of EDA include:
■■ understanding data properties,
■■ finding patterns and relationships in data,
■■ inspecting basic questions and hypotheses,
■■ documenting data distributions and other characteristics, and
■■ planning modeling strategies for the next steps.
Feature selection is a process whereby only pertinent features from the dataset
are selected for ML model training. Selecting fewer features decreases ML model
complexity and training time. Feature engineering is a process of creating new
features by changing or transforming existing features. Model performance heavily
depends on feature selection and engineering.

5.1  Structured Data

Exploratory Data Analysis
For structured data, each data table row contains an observation and each column
contains a feature. EDA can be performed on a single feature (one-­dimension) or on
multiple features (multi-­dimension). For high-­dimension data with many features,
EDA can be facilitated by using a dimension reduction technique, such as principal
components analysis (PCA). Based on the number of dimensions, the exploratory
techniques will vary.
For one-­dimensional data, summary statistics, such as mean, median, quartiles,
ranges, standard deviations, skewness, and kurtosis, of a feature can be computed.
One-­dimension visualization summarizes each feature in the dataset. The basic one-­
dimension exploratory visualizations are as follows:
■■ Histograms
■■ Bar charts
■■ Box plots
■■ Density plots
Histograms represent equal bins of data and their respective frequencies. They can
be used to understand the high-­level distribution of the data. Bar charts summarize
the frequencies of categorical variables. Box plots show the distribution of continuous
data by highlighting the median, quartiles, and outliers of a feature that is normally
distributed. Density plots are another effective way to understand the distribution
of continuous data. Density plots are smoothed histograms and are commonly laid
on top of histograms, as shown in Exhibit 14. This histogram shows a hypothetical
annual salary distribution (in ₤) of entry-­level analyst positions at UK banks. The data
represent a normal distribution with an approximate mean of ₤68,500.
 © CFA Institute. For candidate use only. Not for distribution.
Data Exploration Objectives and Methods 541

In addition to visualization, descriptive statistics are a good means to summarize

data. Central tendency measures as well as minimum and maximum values for con-
tinuous data are useful. Counts and frequencies for categorical data are commonly
employed to gain insight regarding the distribution of possible values.
EDA is not only useful for revealing possible relationships among features or general
trends in the data; it is also beneficial during the feature selection and engineering
stages. These possible relationships and trends in the data may be used to suggest new
features that, when incorporated into a model, may improve model training.

Feature Selection
Structured data consist of features, represented by different columns of data in a table
or matrix. After using EDA to discover relevant patterns in the data, it is essential to
identify and remove unneeded, irrelevant, and redundant features. Basic diagnostic
testing should also be performed on features to identify redundancy, heteroscedas-
ticity, and multi-­collinearity. The objective of the feature selection process is to assist
in identifying significant features that when used in a model retain the important
patterns and complexities of the larger dataset while requiring fewer data overall.
This last point is important since computing power is not free (i.e., explicit costs and
processing time).
Typically, structured data even after the data preparation step can contain features
that do not contribute to the accuracy of an ML model or that negatively affect the
quality of ML training. The most desirable outcome is a parsimonious model with
fewer features that provides the maximum predictive power out-­of-­sample.
Feature selection must not be confused with the data preprocessing steps during
data preparation. Good feature selection requires an understanding of the data and
statistics, and comprehensive EDA must be performed to assist with this step. Data
preprocessing needs clarification only from data administrators and basic intuition
(e.g., salary vs. income) during data preparation.
Feature selection on structured data is a methodical and iterative process. Statistical
measures can be used to assign a score gauging the importance of each feature. The
features can then be ranked using this score and either retained or eliminated from
the dataset. The statistical methods utilized for this task are usually univariate and
consider each feature independently or with regard to the target variable. Methods
include chi-­square test, correlation coefficients, and information-gain measures (i.e.,
R-squared values from regression analysis). All of these statistical methods can be
combined in a manner that uses each method individually on each feature, automatically
performing backward and forward passes over features to improve feature selection.
Prebuilt feature selection functions are available in popular programming languages
used to build and train ML models.
Dimensionality reduction assists in identifying the features in the data that account
for the greatest variance between observations and allows for the processing of a
reduced volume of data. Dimensionality reduction may be implemented to reduce
a large number of features, which helps reduce the memory needed and speed up
learning algorithms. Feature selection is different from dimensionality reduction, but
both methods seek to reduce the number of features in the dataset. The dimension-
ality reduction method creates new combinations of features that are uncorrelated,
whereas feature selection includes and excludes features present in the data without
altering them.

Feature Engineering
After the appropriate features are selected, feature engineering helps further optimize
and improve the features. The success of ML model training depends on how well the
data are presented to the model. The feature engineering process attempts to pro-
duce good features that describe the structures inherent in the dataset. This process
 © CFA Institute. For candidate use only. Not for distribution.
542 Reading 8 ■ Big Data Projects

depends on the context of the project, domain of the data, and nature of the problem.
Structured data are likely to contain quantities, which can be engineered to better
present relevant patterns in the dataset. This action involves engineering an existing
feature into a new feature or decomposing it into multiple features.
For continuous data, a new feature may be created—for example, by taking the
logarithm of the product of two or more features. As another example, when con-
sidering a salary or income feature, it may be important to recognize that different
salary brackets impose a different taxation rate. Domain knowledge can be used to
decompose an income feature into different tax brackets, resulting in a new feature:
“income_above_100k,” with possible values 0 and 1. The value 1 under the new feature
captures the fact that a subject has an annual salary of more than $100,000. By group-
ing subjects into income categories, assumptions about income tax can be made and
utilized in a model that uses the income tax implications of higher and lower salaries
to make financial predictions.
For categorical data, for example, a new feature can be a combination (e.g., sum
or product) of two features or a decomposition of one feature into many. If a single
categorical feature represents education level with five possible values—high school,
associates, bachelor’s, master’s, and doctorate—then these values can be decomposed
into five new features, one for each possible value (e.g., is_highSchool, is_doctorate)
filled with 0s (for false) and 1s (for true). The process in which categorical variables are
converted into binary form (0 or 1) for machine reading is called one hot encoding.
It is one of the most common methods for handling categorical features in text data.
When date-­time is present in the data, such features as “second of the hour,” “hour
of the day,” and “day of the date” can be engineered to capture critical information
about temporal data attributes—which are important, for example, in modeling
trading algorithms.
Feature engineering techniques systemically alter, decompose, or combine existing
features to produce more meaningful features. More meaningful features allow an
ML model to train more swiftly and easily. Different feature engineering strategies
can lead to the generation of dramatically different results from the same ML model.
The impact of feature selection and engineering on ML training is discussed further
in the next section.

5.2  Unstructured Data: Text Exploration

Exploratory Data Analysis
Just like with structured data, it is important to gain insight into existing patterns in
the unstructured data for further analysis. In this section, text data will be discussed.
Text analytics has various applications. The most common applications are text clas-
sification, topic modeling, fraud detection, and sentiment analysis. Text classification
uses supervised ML approaches to classify texts into different classes. Topic modeling
uses unsupervised ML approaches to group the texts in the dataset into topic clusters.
Sentiment analysis predicts sentiment (negative, neutral, or positive) of the texts in a
dataset using both supervised and unsupervised approaches.
Various statistics are used to explore, summarize, and analyze text data. Text data
include a collection of texts (also known as a corpus) that are sequences of tokens.
It is useful to perform EDA of text data by computing on the tokens such basic text
statistics as term frequency (TF), the ratio of the number of times a given token
occurs in all the texts in the dataset to the total number of tokens in the dataset (e.g.,
word associations, average word and sentence length, and word and syllable counts).
Text statistics reveal patterns in the co-­occurrence of words. There are many appli-
cations of text analytics, and necessary text statistics vary according to the context of
the application. Topic modeling is a text data application in which the words that are
 © CFA Institute. For candidate use only. Not for distribution.
Data Exploration Objectives and Methods 545

Feature Engineering
As with structured data, feature engineering can greatly improve ML model training
and remains a combination of art and science. The following are some techniques for
feature engineering, which may overlap with text processing techniques.
1 Numbers: In text processing, numbers are converted into a token, such as “/
number/.” However, numbers can be of different lengths of digits representing
different kinds of numbers, so it may be useful to convert different numbers
into different tokens. For example, numbers with four digits may indicate years,
and numbers with many digits could be an identification number. Four-­digit
numbers can be replaced with “/number4/,” 10-­digit numbers with “/num-
ber10/,” and so forth.
2 N-­grams: Multi-­word patterns that are particularly discriminative can be identi-
fied and their connection kept intact. For example, “market” is a common word
that can be indicative of many subjects or classes; the words “stock market”
are used in a particular context and may be helpful to distinguish general texts
from finance-­related texts. Here, a bigram would be useful as it treats the two
adjacent words as a single token (e.g., stock_market).
3 Name entity recognition (NER): NER is an extensive procedure available as a
library or package in many programming languages. The name entity rec-
ognition algorithm analyzes the individual tokens and their surrounding
semantics while referring to its dictionary to tag an object class to the token.
Exhibit 19 shows the NER tags of the text “CFA Institute was formed in 1947
and is headquartered in Virginia.” Additional object classes are, for example,
MONEY, TIME, and PERCENT, which are not present in the example text. The
NER tags, when applicable, can be used as features for ML model training for
better model performance. NER tags can also help identify critical tokens on
which such operations as lowercasing and stemming then can be avoided (e.g.,
Institute here refers to an organization rather than a verb). Such techniques
make the features more discriminative.

Exhibit 19  Name Entity Recognition and Parts of Speech (POS) on Example

Text
Token NER Tag POS Tag POS Description

CFA ORGANIZATION NNP Proper noun

Institute ORGANIZATION NNP Proper noun
was VBD Verb, past tense
formed VBN Verb, past participle
in IN Preposition
1947 DATE CD Cardinal number
and CC Coordinating conjunction
is VBZ Verb, 3rd person singular present
headquartered VBN Verb, past participle
(continued)
 © CFA Institute. For candidate use only. Not for distribution.
546 Reading 8 ■ Big Data Projects

Exhibit 19  (Continued)

Token NER Tag POS Tag POS Description

in IN Preposition
Virginia LOCATION NNP Proper noun

4 Parts of speech (POS): Similar to NER, parts of speech uses language structure
and dictionaries to tag every token in the text with a corresponding part of
speech. Some common POS tags are noun, verb, adjective, and proper noun.
Exhibit 19 shows the POS tags and descriptions of tags for the example text.
POS tags can be used as features for ML model training and to identify the
number of tokens that belong to each POS tag. If a given text contains many
proper nouns, it means that it may be related to people and organizations and
may be a business topic. POS tags can be useful for separating verbs and nouns
for text analytics. For example, the word “market” can be a verb when used as
“to market …” or noun when used as “in the market.” Differentiating such tokens
can help further clarify the meaning of the text. The use of “market” as a verb
could indicate that the text relates to the topic of marketing and might discuss
marketing a product or service. The use of “market” as a noun could suggest
that the text relates to a physical or stock market and might discuss stock
trading. Also for POS tagging, such compound nouns as “CFA Institute” can
be treated as a single token. POS tagging can be performed using libraries or
packages in programming languages.
In addition, many more creative techniques convey text information in a struc-
tured way to the ML training process. The goal of feature engineering is to maintain
the semantic essence of the text while simplifying and converting it into structured
data for ML.

EXAMPLE 4 

Data Exploration
Paul Wang’s analytics team at LendALot Corporation has completed its initial
data preparation and wrangling related to their creditworthiness classification
ML model building efforts. As a next step, Wang has asked one of the team
members, Eric Kim, to examine the available structured data sources to see what
types of exploratory data analysis might make sense. Kim has been tasked with
reporting to the team on high-­level patterns and trends in the data and which
variables seem interesting. Greater situational awareness about the data can
inform the team’s decisions regarding model training and whether (and how) to
incorporate textual big data in conjunction with the structured data inputs. Use
the following sample of columns and rows Kim pulled for manual examination
to answer the next questions.
 © CFA Institute. For candidate use only. Not for distribution.
Model Training 549

Based on the information given, address the following questions.

1 Describe three textual feature representations that Lee and Kim should
consider for their text data.
2 Describe a rationale for adopting each of the three textual feature repre-
sentations identified in Question 1.

Solution 1:
Lee and Kim should consider bag-­of-­words (BOW), n-­grams, and parts-­of-­speech
(POS) as key textual feature representations for their text data. Conversely, name
entity recognition (NER) might not be as applicable in this context because the
data on prospective borrowers does not include any explicit references to people,
locations, dates, or organizations.
Solution 2:
All three textual feature representations have the potential to add value.
Bag-­of-­words (BOW) is typically applicable in most contexts involving text
features derived from languages where token boundaries are explicitly present
(e.g., English) or can be inferred through processing (e.g., a different language,
such as Spanish). BOW is generally the best starting point for most projects
exploring text feature representations.
N-­grams, representations of word or token sequences, are also applicable.
N-­grams can offer invaluable contextual information that can complement and
enrich a BOW. In this specific credit-­worthiness context, we examine the BOW
token “worked.” It appears three times (rows 5–7), twice in no-­default loan texts
and once in a defaulted loan text. This finding suggests that “worked” is being
used to refer to the borrower’s work ethic and may be a good predictor of credit
worthiness. Digging deeper and looking at several trigrams (i.e., three-­token
sequences) involving “worked,” we see that “have_worked_hard” appears in the
two no-­default loan related texts (referring to borrower accomplishments and
plans) and “had_worked_harder” appears in the defaulted loan text (referring to
what could have been done). This example illustrates how n-­grams can provide
richer contextualization capabilities for the creditworthiness prediction ML
models.
Parts-­of-­speech tags can add value because they identify the composition of
the texts. For example, POS provides information on whether the prospective
borrowers are including many action words (verbs) or descriptors (adjectives)
and whether this is being done differently in instances of no-­default versus
instances of defaulted loans.

MODEL TRAINING
Machine learning model training is a systematic, iterative, and recursive process. The
6
number of iterations required to reach optimum results depends on:
■■ the nature of the problem and input data and
■■ the level of model performance needed for practical application.
Machine learning models combine multiple principles and operations to provide
predictions. As seen in the last two sections, typical ML model building requires
data preparation and wrangling (cleansing and preprocessing) and data exploration
(exploratory data analysis as well as feature selection and engineering). In addition,
 © CFA Institute. For candidate use only. Not for distribution.
550 Reading 8 ■ Big Data Projects

domain knowledge related to the nature of the data is required for good model build-
ing and training. For instance, knowledge of investment management and securities
trading is important when using financial data to train a model for predicting costs
of trading stocks. It is crucial for ML engineers and domain experts to work together
in building and training robust ML models.
The three tasks of ML model training are method selection, performance evaluation,
and tuning. Exhibit 20 outlines model training and its three component tasks. Method
selection is the art and science of deciding which ML method(s) to incorporate and
is guided by such considerations as the classification task, type of data, and size of
data. Performance evaluation entails using an array of complementary techniques and
measures to quantify and understand a model’s performance. Tuning is the process of
undertaking decisions and actions to improve model performance. These steps may be
repeated multiple times until the desired level of ML model performance is attained.
Although no standard rulebook for training an ML model exists, having a fundamental
understanding of domain-­specific training data and ML algorithm principles plays a
vital role in good model training.

Exhibit 20  Model Training Stage

Model Training
Data Data
Collection/ Preparation Data
and Method Performance Results
Curation Exploration Tuning
Wrangling Selection Evaluation

Before training a model, it is important to state the problem, define objectives,

identify useful data points, and conceptualize the model. Conceptualization is like
a blueprint on a drawing board, a modifiable plan that is necessary to initiate the
model training process. Because modeling is an iterative process, many changes and
refinements will be made to the model plan as the process evolves.

6.1  Structured and Unstructured Data

The ML model training process for structured and unstructured data is typically the
same. Most ML models are intended to train on structured data, so unstructured data
in the data preparation stage are processed and organized into a structured format.
The systematic processing of unstructured text data so that they can be structured
in the form of a data matrix has been previously covered. Similarly, other forms of
unstructured data can also be prepared and formed into data matrixes or tables for
ML training.
The fundamental idea of ML model training is fitting a system of rules on a training
dataset to reveal a pattern in the data. In other words, fitting describes the degree to
which (or how well) an ML model can be generalized to new data. A good model fit
results in good model performance and can be validated using new data outside of
the training dataset (i.e., out-­of-­sample). Exhibit 21 shows model decision boundar-
ies in three possible model fitting scenarios for a classification task comprising two
different classes of data (i.e., circles and triangles). The model on the left is underfit;
it does not fit the training data well enough since it results in four misclassification
errors (three circles and one triangle). Although the center model that generates the
“S”-shaped line has the best accuracy (no errors) on the training data, it is overfit
(i.e., fits the training data too well) and thus unlikely to perform well on future test
 © CFA Institute. For candidate use only. Not for distribution.
Model Training 553

Exhibit 22 depicts the idea of undersampling of the majority class and oversampling
of the minority class. In practice, the choice of whether to undersample or oversample
depends on the specific problem context. Advanced techniques can also reproduce
synthetic observations from the existing data, and the new observations can be added
to the dataset to balance the minority class.

Exhibit 22  Undersampling and Oversampling

Undersampling Majority Class (”0” class)

“1” Class “0” Class “0” Class “1” Class

Original Dataset Final Dataset

Oversampling Minority Class (”1” class)

“0” Class “1” Class “1” Class “0” Class

Original Dataset Final Dataset

Performance Evaluation
It is important to measure the model training performance or goodness of fit for vali-
dation of the model. We shall cover several techniques to measure model performance
that are well suited specifically for binary classification models.
1 Error analysis. For classification problems, error analysis involves computing
four basic evaluation metrics: true positive (TP), false positive (FP), true nega-
tive (TN), and false negative (FN) metrics. FP is also called a Type I error, and
FN is also called a Type II error. Exhibit 23 shows a confusion matrix, a grid
that is used to summarize values of these four metrics.
 © CFA Institute. For candidate use only. Not for distribution.
554 Reading 8 ■ Big Data Projects

Exhibit 23  Confusion Matrix for Error Analysis

Actual Training Labels
Class “1” Class “0”

Class “1” True Positives (TP) False Positives (FP)

Type I Error

Predicted Results

Class “0” False Negatives (FN) True Negatives (TN)

Type II Error

Additional metrics, such as precision and recall, can be computed. Assume

in the following explanation that Class “0” is “not defective” and Class “1” is
“defective.” Precision is the ratio of correctly predicted positive classes to all
predicted positive classes. Precision is useful in situations where the cost of FP,
or Type I error, is high—or example, when an expensive product fails quality
inspection (predicted Class “1”) and is scrapped, but it is actually perfectly good
(actual Class “0”). Recall (also known as sensitivity) is the ratio of correctly pre-
dicted positive classes to all actual positive classes. Recall is useful in situations
where the cost of FN or Type II error is high—for example, when an expensive
product passes quality inspection (predicted Class “0”) and is sent to the valued
customer, but it is actually quite defective (actual Class “1”). The formulas for
precision and recall are:
Precision (P) = TP/(TP + FP). (3)
Recall (R) = TP/(TP + FN). (4)

Trading off precision and recall is subject to business decisions and model
application. Therefore, additional evaluation metrics that provide the overall
performance of the model are generally used. The two overall performance
metrics are accuracy and F1 score. Accuracy is the percentage of correctly
predicted classes out of total predictions. F1 score is the harmonic mean of
precision and recall. F1 score is more appropriate (than accuracy) when unequal
class distribution is in the dataset and it is necessary to measure the equilib-
rium of precision and recall. High scores on both of these metrics suggest good
model performance. The formulas for accuracy and F1 score are as follows:
Accuracy = (TP + TN)/(TP + FP + TN + FN). (5)
F1 score = (2 * P * R)/(P + R). (6)

Exhibit 24 illustrates computations of model evaluation metrics and perfor-

mance scores on a sample dataset.
 © CFA Institute. For candidate use only. Not for distribution.
Model Training 557

in the training data. Variance error is high when the model is overly complicated
and memorizes the training data so much that it will likely perform poorly on new
data. It is not possible to completely eliminate both types of errors. However, both
errors can be minimized so the total aggregate error (bias error + variance error) is
at a minimum. The bias–variance trade-­off is critical to finding an optimum balance
where a model neither underfits nor overfits.
1 Parameters are critical for a model and are dependent on the training data.
Parameters are learned from the training data as part of the training process
by an optimization technique. Examples of parameters include coefficients in
regression, weights in NN, and support vectors in SVM.
2 Hyperparameters are used for estimating model parameters and are not depen-
dent on the training data. Examples of hyperparameters include the regulariza-
tion term (λ) in supervised models, activation function and number of hidden
layers in NN, number of trees and tree depth in ensemble methods, k in k-­near-
est neighbor classification and k-­means clustering, and p-­threshold in logistic
regression. Hyperparameters are manually set and tuned.
For example, if a researcher is using a logistic regression model to classify sentences
from financial statements into positive or negative stock sentiment, the initial cutoff
point for the trained model might be a p-­threshold of 0.50 (50%). Therefore, any sen-
tence for which the model produces a probability >50% is classified as having positive
sentiment. The researcher can create a confusion matrix from the classification results
(of running the CV dataset) to determine such model performance metrics as accuracy
and F1 score. Next, the researcher can vary the logistic regression’s p-­threshold—say
to 0.55 (55%), 0.60 (60%), or even 0.65 (65%)—and then re-­run the CV set, create new
confusion matrixes from the new classification results, and compare accuracy and F1
scores. Ultimately, the researcher would select the logistic regression model with a
p-­threshold value that produces classification results generating the highest accuracy
and F1 scores. Note that the process just outlined will be demonstrated in Section 7.
There is no general formula to estimate hyperparameters. Thus, tuning heuristics
and such techniques as grid search are used to obtain the optimum values of hyper-
parameters. Grid search is a method of systematically training an ML model by using
various combinations of hyperparameter values, cross validating each model, and
determining which combination of hyperparameter values ensures the best model
performance. The model is trained using different combinations of hyperparameter
values until the optimum set of values are found. Optimum values must result in sim-
ilar performance of the model on training and CV datasets, meaning that the training
error and CV error are close. This ensures that the model can be generalized to test
data or to new data and thus is less likely to overfit. The plot of training errors for
each value of a hyperparameter (i.e., changing model complexity) is called a fitting
curve. Fitting curves provide visual insight on the model’s performance (for the given
hyperparameter and level of model complexity) on the training and CV datasets and
are visually helpful to tune hyperparameters. Exhibit 26 shows the bias–variance error
trade-­off by plotting a generic fitting curve for a regularization hyperparameter (λ).
 © CFA Institute. For candidate use only. Not for distribution.
558 Reading 8 ■ Big Data Projects

Exhibit 26  Fitting Curve for Regularization Hyperparameter (λ)

Large Error
High Variance High Bias

Errorcv

Errortrain

Errorcv >>Errortrain
Error
Overfitting
Underfitting

Optimum Regularization
Small Error
Slight Regularization Large Regularization
Lambda (˜ )

Slight regularization lightly penalizes model complexity, thereby allowing most or

all of the features to be included in the model and thus potentially enabling the model
to “memorize” the data. Typically with no or slight regularization, the prediction error
on the training dataset is small while the prediction error on the CV dataset is signifi-
cantly larger. This difference in error is variance. High variance error, which typically
results from too many features and model complexity, results in model overfitting.
When high variance error and low bias error exist, the model performs well on the
training dataset but generates many FP and FN errors on the CV dataset; in other
words, the model is overfitted and does not generalize to new data well.
Large regularization excessively penalizes model complexity, thereby allowing too
few of the features to be included in the model and causing the model to learn less
from the data. The model may lack the necessary predictor variables and complexity
needed to discern underlying patterns in the data. Typically with large regularization,
the prediction errors on the training and CV datasets are both large. Large prediction
errors on the training dataset indicate high bias, and high bias error results from model
underfitting. When high bias error exists, the model does not perform well on either
training or CV datasets because it is typically lacking important predictor variables.
Optimum regularization minimizes both variance and bias errors in a balanced
fashion. It penalizes model complexity just enough so that only the most important
features are included in the model. This process prevents the model from memoriz-
ing the data while enabling the model to learn enough from the data to distinguish
important patterns. This results in prediction errors in both training and CV datasets
that are similar and also minimal. The range of optimum regularization values can be
found heuristically using such techniques as grid search.
If high bias or variance exists after the tuning of hyperparameters, either a larger
number of training examples (instances) may be needed or the number of features
included in the model may need to be decreased (in the case of high variance) or
increased (in the case of high bias). The model then needs to be re-­trained and re-­tuned
using the new training dataset. In the case of a complex model, where a large model
 © CFA Institute. For candidate use only. Not for distribution.
Financial Forecasting Project: Classifying and Predicting Sentiment for Stocks 561

The raw text contains punctuations, numbers, and white spaces that may not be
necessary for model training. Text cleansing involves removing, or incorporating
appropriate substitutions for, potentially extraneous information present in the text.
Operations to remove html tags are unnecessary because none are present in the text
Punctuations: Before stripping out punctuations, percentage and dollar symbols are
substituted with word annotations to retain their essence in the financial texts. Such
word annotation substitutions convey that percentage and currency-­related tokens were
involved in the text. As the sentences have already been identified within and extracted
from the source text, punctuation helpful for identifying discrete sentences—such as
periods, semi-­colons, and commas—are removed. Some special characters, such as
“+” and “©,” are also removed. It is a good practice to implement word annotation
substitutions before removing the rest of the punctuations.
Numbers: Numerical values of numbers in the text have no significant utility for
sentiment prediction in this project because sentiment primarily depends on the words
in a sentence. Here is an example sentence: “Ragutis, which is based in Lithuania's
second-­largest city, Kaunas, boosted its sales last year 22.3  percent to 36.4  million
litas.” The word “boosted” implies that there was growth in sales, so analysis of this
sentiment does not need to rely on interpretation of numerical text data. Sentiment
analysis typically does not involve extracting, interpreting, and calculating relevant
numbers but instead seeks to understand the context in which the numbers are used.
Other commonly occurring numbers are dates and years, which are also not required
to predict sentence sentiment. Thus, all numbers present in the text are removed for
this financial sentiment project. However, prior to removing numbers, abbreviations
representing orders of magnitude, such as million (commonly represented by “m,”
“mln,” or “mn”), billion, or trillion, are replaced with the complete word. Retaining
these orders of magnitude-­identifying words in the text preserves the original text
meaning and can be useful in predicting sentence sentiment.
Whitespaces: White spaces are present in the raw text. Additional white spaces
occur after performing the above operations to remove extraneous characters. The
white spaces must be removed to keep the text intact. Exhibit 29 shows the sample
text after cleansing. The cleansed text is free of punctuations and numbers, with
useful substitutions.

Exhibit 29  Ten Sample Rows After Cleansing Process

Sentence Sentiment

Profit before taxes amounted to EUR million down from EUR million a year ago negative
Profit before taxes decreased by percentSign to EUR million in the first nine months of compared to negative
EUR million a year earlier
Profit before taxes decreased to EUR million from EUR million the year before negative
Profit before taxes was EUR million down from EUR million negative
The companys profit before taxes fell to EUR million in the third quarter of compared to EUR million negative
in the corresponding period in
In August October the companys result before taxes totalled EUR million up from EUR million in the positive
corresponding period in
Finnish Bore that is owned by the Rettig family has grown recently through the acquisition of smaller positive
shipping companies
The plan is estimated to generate some EUR million USD million in cost savings on an annual basis positive
(continued)
 © CFA Institute. For candidate use only. Not for distribution.
562 Reading 8 ■ Big Data Projects

Exhibit 29  (Continued)

Sentence Sentiment
Finnish pharmaceuticals company Orion reports profit before taxes of EUR million in the third quar- positive
ter of up from EUR million in the corresponding period in
Finnish Sampo Bank of Danish Danske Bank group reports profit before taxes of EUR million in up positive
from EUR million in

Text Wrangling (Preprocessing)

The cleansed text needs to be normalized using the following normalization procedures:
1 Lowercasing of all text to consolidate duplicate words (example, “THE,” “The,”
and “the”).
2 Stop words are not removed because some stop words (e.g., not, more, very, and
few) carry significant meaning in the financial texts that is useful for sentiment
prediction. Some stop words, such as articles (a, an, the), may be removed.
Nevertheless, to avoid confusion no words are removed at this point. This issue
will be revisited during the data exploration stage, which will carefully examine
the text using frequency analysis and find custom stop words (common words)
for these particular text data.
3 Stemming, the converting of inflected forms of a word into its base word (stem),
is performed on the text as it is simple to perform and is appropriate for train-
ing an ML model for sentiment prediction.
White spaces are stripped after performing these operations. As part of text nor-
malization, different currency abbreviations, such as EUR and USD, can be converted
into a single token, such as “currencysign.” As we are dealing with financial domain
text, the earlier substitution of dollarsign can be replaced with currencysign as well.
This step will remove tokens that are different but redundant in nature while main-
taining their meaning. Through careful examination of the text and use of domain
knowledge, similar substitutions of redundant tokens can be performed. Exhibit 30
shows how the sample text appears after normalization.

Exhibit 30  Ten Sample Rows After Normalization Process

Sentence Sentiment

profit befor tax amount to currencysign million down from currencysign million a year ago negative
profit befor tax decreas by percentsign to currencysign million in the first nine month of compar to negative
currencysign million a year earlier
profit before tax decreas to currencysign million from currencysign million the year befor negative
profit befor tax was currencysign million down from currencysign million negative
the compani profit befor tax fell to currencysign million in the third quarter of compar to currency- negative
sign million in the correspond period in
in august octob the compani result befor tax total currencysign million up from currencysign million positive
in the correspond period in
finnish bore that is own by the rettig famili has grown recent through the acquisit of smaller shipping positive
company
the plan is estim to generat some currencysign million currencysign million in cost save on an annual positive
basi
 © CFA Institute. For candidate use only. Not for distribution.
Financial Forecasting Project: Classifying and Predicting Sentiment for Stocks 565

Exhibit 34 shows a histogram of sentence length distribution. Sentence length

is defined as the number of characters, including spaces, in a sentence. The longest
sentence has 273 characters; the shortest sentence has 26 characters; and the average
number of characters is about 120 (indicated by the vertical line). Although this dis-
tribution does not have any direct impact on model training, this histogram visually
demonstrates the range of sentence lengths and helps identify any extremely long
or short sentences. This histogram does not appear unusual, so no outlier sentences
need to be removed.

Exhibit 34  Histogram of Sentence Lengths with Mean Sentence Length

Frequency
400

300

200

100

0
0 50 100 150 200 250 300
Number of Characters in a Sentence

Word clouds are a convenient method of visualizing the text data because they
enable rapid comprehension of a large number of tokens and their corresponding
weights. Exhibit 35 shows a word cloud for all the sentences in the corpus. The font
sizes of the words are proportionate to the number of occurrences of each word in the
corpus. Similarly, Exhibit 36 shows the word cloud divided into two halves: one half
representing negative sentiment class sentences (upper half ); one half representing
positive sentiment class sentences (lower half ). Notably, some highly discriminative
stems and words, such as “decreas” and “down” in the negative half and “increas” and
“rose” in the positive half, are present. The feature selection process will eliminate
common words and highlight useful words for better model training.
 © CFA Institute. For candidate use only. Not for distribution.
566 Reading 8 ■ Big Data Projects

Exhibit 35  Word Cloud of Entire Corpus

Exhibit 36  Word Cloud Divided by Two Sub-­Groups of the Corpus

Feature Selection
Exploratory data analysis revealed the most frequent tokens in the texts that could
potentially add noise to this ML model training process. In addition to common tokens,
many rarely occurring tokens, often proper nouns (i.e., names), are not informative
for understanding the sentiment of the sentence. Further analyses must be conducted
to decide which words to eliminate. Feature selection for text data involves keeping
the useful tokens in the BOW that are informative and help to discriminate different
classes of texts—those with positive sentiment and those with negative sentiment. At
this point, a total of 44,151 non-­unique tokens are in the 2,180 sentences.
Frequency analysis on the processed text data helps in filtering unnecessary
tokens (or features) by quantifying how important tokens are in a sentence and in the
corpus as a whole. Term frequency (TF) at the corpus level—also known as collection
frequency (CF)—is the number of times a given word appears in the whole corpus
 © CFA Institute. For candidate use only. Not for distribution.
Financial Forecasting Project: Classifying and Predicting Sentiment for Stocks 569

For example, TF at the sentence level for the word “the” in sentences num-
ber 701 and 223 is calculated as 6/39 = 0.1538462 and 5/37 = 0.1351351,
respectively.
8 DF (Document Frequency): Defined as the number of documents (i.e., sentences)
that contain a given word divided by the total number of sentences (here,
2,180). Document frequency is important since words frequently occurring
across sentences provide no differentiating information in each sentence. The
following equation can be used to compute DF:
DF = SentenceCountWithWord/Total number of sentences. (12)

For example, DF of the word “the” is 1,453/2,180 = 0.6665138; so, 66.7% of the
sentences contain the word “the.” A high DF indicates high word frequency in
the text.
9 IDF (Inverse Document Frequency): A relative measure of how unique a term
is across the entire corpus. Its meaning is not directly related to the size of the
corpus. The following equation can be used to compute IDF:
IDF = log(1/DF). (13)

For example, IDF of the word “the” is log(1/0.6665138) = 0.4056945. A low IDF
indicates high word frequency in the text.
10 TF–IDF: To get a complete representation of the value of each word, TF at
the sentence level is multiplied by the IDF of a word across the entire dataset.
Higher TF–IDF values indicate words that appear more frequently within a
smaller number of documents. This signifies relatively more unique terms that
are important. Conversely, a low TF–IDF value indicates terms that appear in
many documents. TF–IDF values can be useful in measuring the key terms
across a compilation of documents and can serve as word feature values for
training an ML model. The following equation can be used to compute TF–IDF:
TF–IDF = TF × IDF. (14)

For example, TF–IDF of the token “of ” is calculated as 0.1666667 × 0.7954543 =

0.13257571.
Similarly, Exhibit 40 shows high TF–IDF words for the text data before the removal
of custom stop words.

Exhibit 40  Sample Output of High TF–IDF Words

SentenceNo TotalWordsInSentence Word TotalWordCount WordCountInSentence SentenceCountWithWord TF DF IDF TFIDF

<int> <int> <chr> <int> <int> <int> <dbl> <dbl> <dbl> <dbl>
28 7 risen 3 1 3 0.1428571 0.0013761468 6.588468 0.9412097
830 7 diminish 2 1 2 0.1428571 0.0009174312 6.993933 0.9991333
1368 9 great 4 1 4 0.1111111 0.0018348624 6.300786 0.7000873
1848 8 injuri 1 1 1 0.1250000 0.0004587156 7.687080 0.9608850
1912 7 cheaper 1 1 1 0.1428571 0.0004587156 7.687080 1.0981543
1952 6 argument 1 1 1 0.1666667 0.0004587156 7.687080 1.2811800

TF or TF–IDF values are placed at the intersection of sentences (rows) and terms
(columns) of the document term matrix. For this project, TF values are used for the
DTM as the texts are sentences rather than paragraphs or other larger bodies of
text. TF–IDF values vary by the number of documents in the dataset; therefore, the
 © CFA Institute. For candidate use only. Not for distribution.
570 Reading 8 ■ Big Data Projects

model performance can vary when applied to a dataset with just a few documents.
In addition to removing custom stop words and sparse terms, single character letters
are also eliminated because they do not add any value to the sentiment significance.

Feature Engineering
N-­grams are used as a feature engineering process in this project. Use of n-­grams
helps to understand the sentiment of a sentence as a whole. As mentioned previously,
the objective of this project is to predict sentiment class (positive and negative) from
financial texts. Both unigram and bigrams are implemented, and the BOW is created
from them. Bigram tokens are helpful for keeping negations intact in the text, which
is vital for sentiment prediction. For example, the tokens “not” and “good” or “no”
and “longer” can be formed into single tokens, now bigrams, such as “not_good”
and “no_longer.” These and similar tokens can be useful during ML model training
and can improve model performance. Exhibit 41 shows a sample of 100 words from
the BOW containing both unigram and bigram tokens after removal of custom stop
words, sparse terms, and single characters. Note that the BOW contains such tokens
as increas, loss, loss_prior, oper_rose, tax_loss, and sale_increas. Such tokens are
informative about the embedded sentiment in the texts and are useful for training
an ML model. The corresponding word frequency measures for the document term
matrix are computed based on this new BOW.

Exhibit 41  One-­Hundred Sample Tokens from Final BOW of Entire Corpus

“last” “last_quarter” “quarter” “quarter_componenta” “componenta”

“componenta_sale” “sale” “sale_doubl” “doubl” “doubl_same”
“same” “same_period” “period” “period_earlier” “earlier”
“earlier_while” “while” “while_move” “move” “move_zero”
“zero” “zero_pre” “pre” “pre_tax” “tax”
“tax_pre” “tax_loss” “loss” “third” “third_quarter”
“quarter_sale” “sale_increas” “increas” “increas_by” “by”
“by_percentsign” “percentsign” “percentsign_oper” “oper” “oper_by”
“oper_rose” “rose” “rose_correspond” “correspond” “correspond_period”
“period_repres” “repres” “repres_percentsign” “percentsign_sale” “oper_total”
“total” “total_up” “up” “up_repres” “finnish”
“finnish_talentum” “talentum” “talentum_report” “report” “report_oper”
“oper_increas” “increas_sale” “sale_total” “cloth” “cloth_retail”
“retail” “retail_chain” “chain” “chain_sepp” “sepp”
“sepp_ls” “ls” “ls_sale” “consolid” “consolid_sale”
“incres_percentsign” “percentsign_reach” “reach” “reach_while” “while_oper”
“oper_amount” “amount” “amount_compar” “compar” “compar_loss”
“loss_prior” “prior” “prior_period” “foundri” “foundri_divis”
“divis” “divis_report” “report_sale” “percentsign_correspond” “period_sale”
“sale_machin” “machin” “machin_shop” “shop” “shop_divis”

EXAMPLE 6 

Calculating and Interpreting Term Frequency Measures

Data scientists Jack and Jill are using financial text data to develop sentiment
indicators for forecasting future stock price movements. They have assembled
a BOW from the corpus of text being examined and have pulled the following
abbreviated term frequency measures tables.
 © CFA Institute. For candidate use only. Not for distribution.
Financial Forecasting Project: Classifying and Predicting Sentiment for Stocks 573

Similarly, the DTMs for the CV set and the test set are built using tokens from the
final training BOW for tuning, validating, and testing of the model. To be clear, the
final BOW from the training corpus is used for building DTMs across all the splits
because the model has been trained on that final BOW. Thus, the columns (think,
features) of all three DTMs are the same, but the number of rows varies because a
different number of sentences are in each split. The DTMs are filled with resultant
term frequency values calculated using sentences in the corpuses of the respective
splits—sentences from the CV set corpus and sentences from the test set corpus.
Exhibit 42 tabulates the summary of dimensions of the data splits and their uses in
the model training process. As mentioned, the columns of DTMs for the splits are
the same, equal to the number of unique tokens (i.e., features) from the final training
corpus BOW, which is 9,188. Note that this number of unique tokens (9,188) differs
from that in the master corpus (11,501) based on the sentences that are included in
the training corpus after the random sampling.

Exhibit 42  Summary of the Three Data Splits

Number of
Corpus Split % Sentences DTM Dimensions Purpose

Master 100% 2180 2180 × 11501 Used for data exploration

Training 60% 1309 1309 × 9188 Used for ML model training
CV 20% 435 435 × 9188 Used for tuning and validating the trained model
Test 20% 436 436 × 9188 Used for testing the trained, tuned, and validated
model

Method Selection
Alternative ML methods, including SVM, decision trees, and logistic regression, were
examined because these techniques are all considered potentially suitable for this
particular task (i.e., supervised learning), type of data (i.e., text), and size of data (i.e.,
wider data with many potential variables). The SVM and logistic regression methods
appeared to offer better performance than decision trees. For brevity, we discuss
logistic regression in the remainder of the chapter. Logistic regression was used to
train the model, using the training corpus DTM containing 1,309 sentences. As a
reminder, in this project texts are the sentences and the classifications are positive
and negative sentiment classes (labeled 1 and 0, respectively). The tokens are feature
variables, and the sentiment class is the target variable. Text data typically contain
thousands of tokens. These result in sparse DTMs because each column represents
a token feature and the values are mostly zeros (i.e., not all the tokens are present
in every text). Logistic regression can deal with such sparse training data because
the regression coefficients will be close to zero for tokens that are not present in a
significant number of sentences. This allows the model to ignore a large number of
minimally useful features. Regularization further helps lower the coefficients when
the features rarely occur and do not contribute to the model training.
Logistic regression is applied on the final training DTM for model training. As
this method uses maximum likelihood estimation, the output of the logistic model is
a probability value ranging from 0 to 1. However, because the target variable is binary,
coefficients from the logistic regression model are not directly used to predict the
value of the target variable. Rather, a mathematical function uses the logistic regression
coefficient (β) to calculate probability (p) of sentences having positive sentiment (y =
 © CFA Institute. For candidate use only. Not for distribution.
574 Reading 8 ■ Big Data Projects

1).3 If p for a sentence is 0.90, there is a 90% likelihood that the sentence has positive
sentiment. Theoretically, the sentences with p > 0.50 likely have positive sentiment.
Because this is not always true in practice, however, it is important to find an ideal
threshold value of p. We elaborate on this point in a subsequent example. The thresh-
old value is a cutoff point for p values, and the ideal threshold p value is influenced
by the dataset and model training. When the p values (i.e., probability of sentences
having positive sentiment) of sentences are above this ideal threshold p value, then
the sentences are highly likely to have positive sentiment (y = 1). The ideal threshold
p value is estimated heuristically using performance metrics and ROC curves, as will
be demonstrated shortly.

Performance Evaluation and Tuning

The trained ML model is used to predict the sentiments of the sentences in the training
and CV DTMs. Exhibit 43 displays the ROC curves for the training (Panel A) and CV
(Panel B) data. Remember that the x-­axis is false positive rate, FP/(TN + FP), and the
y-­axis is true positive rate, TP/(TP + FN). As the model is trained using the training
DTM, it clearly performs well on the same training data (so there is no concern about
underfitting) but does not perform as well on the CV data. This is apparent as the ROC
curves are significantly different between the training and CV datasets. The AUC is
96.5% on training data and 86.2% on CV data. This finding suggests that the model
performs comparatively poorly (with a higher rate of error or misclassification) on
the CV data when compared to training data. Thus, the implication is that the model
is overfitted.

Exhibit 43  ROC Curves of Model Results for Training and CV Data Before
Regularization
A. ROC Curve for Training Data B. ROC Curve for CV Data
True Positive Rate True Positive Rate

1.0 1.0

0.8 0.8

0.6 0.6
AUC = AUC =
96.5% 86.2%
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
False Positive Rate False Positive Rate

As the model is overfitted, least absolute shrinkage and selection operator (LASSO)
regularization is applied to the logistic regression. LASSO regularization penalizes the
coefficients of the logistic regression to prevent overfitting of the model. The penal-
ized regression will select the tokens (features) that have statistically significant (i.e.,
non-­zero) coefficients and that contribute to the model fit; LASSO does this while
disregarding the other tokens. Exhibit 44 shows the ROC curves for the new model

1
3  This mathematical function is an exponential function of the form: P (y = 1) =
1 + exp ( 0 1 1 2 2
− β + β x + β x ++ βn xn )

where the βs are the logistic regression coefficients.

 © CFA Institute. For candidate use only. Not for distribution.
Financial Forecasting Project: Classifying and Predicting Sentiment for Stocks 577

Exhibit 46  Threshold Values Versus Overall Performance Measures

F1 Score Accuracy

0.9
0.8
0.8
0.6 0.7

0.4 0.6

0.5
0.2
0.4
0 0
0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
Threshold Value Threshold Value

Exhibit 47  Performance Measures of the Model for a Series of Threshold

Values
Threshold Precision Recall F1 Accuracy

0.45 0.8750000 0.986254296 0.927302100 0.8965517

0.46 0.8827160 0.982817869 0.930081301 0.9011494
0.47 0.8827160 0.982817869 0.930081301 0.9011494
0.48 0.8819876 0.975945017 0.926590538 0.8965517
0.49 0.8819876 0.975945017 0.926590538 0.8965517
0.50 0.8819876 0.975945017 0.926590538 0.8965517
0.51 0.8819876 0.975945017 0.926590538 0.8965517
0.52 0.8819876 0.975945017 0.926590538 0.8965517
0.53 0.8902821 0.975945017 0.931147541 0.9034483
0.54 0.8930818 0.975945017 0.932676519 0.9057471
0.55 0.8930818 0.975945017 0.932676519 0.9057471
0.56 0.8958991 0.975945017 0.934210526 0.9080460
0.57 0.8958991 0.975945017 0.934210526 0.9080460
0.58 0.8958991 9.975945017 0.934210526 0.9080460
0.59 0.9015873 0.975945017 0.937293729 0.9126437
0.60 0.9044586 0.975945017 0.938842975 0.9149425
0.61 0.9044586 0.975945017 0.938842975 0.9149425
0.62 0.9044586 0.975945017 0.938842975 0.9149425
0.63 0.9041534 0.972508591 0.937086093 0.9126437
0.64 0.9041534 0.972508591 0.937086093 0.9126537
0.65 0.9041534 0.972508591 0.937086093 0.9126437
0.66 0.9035370 0.965635739 0.933554817 0.9080460
0.67 0.9035370 0.965635739 0.933554817 0.9080460
0.68 0.9064516 0.965635739 0.935108153 0.9103448
0.69 0.9064516 0.965635739 0.935108153 0.9103448
0.70 0.9061489 0.962199313 0.933333333 0.9080460
0.71 0.9061489 0.962199313 0.933333333 0.9080460
(continued)
 © CFA Institute. For candidate use only. Not for distribution.
578 Reading 8 ■ Big Data Projects

Exhibit 47  (Continued)

Threshold Precision Recall F1 Accuracy

0.72 0.9090909 0.962199313 0.934891486 0.9103448
0.73 0.9090909 0.962199313 0.934891486 0.9103448
0.74 0.9078947 0.948453608 0.927731092 0.9011494
0.75 0.9072848 0.941580756 0.924114671 0.8965517

* The shaded row shows the selected threshold p value (0.60) and the performance metrics for the
selected model.

Finally, the confusion matrix using the ideal threshold p value of 0.60 is constructed
to observe the performance of the final model. When target value p > 0.60, the predic-
tion is assumed to be y = 1 (indicating positive sentiment); otherwise, the prediction
is assumed to be y = 0 (negative sentiment). The confusion matrix for the CV data is
shown in Exhibit 48. It is clear that the model performance metrics have improved
in the final model compared to the earliest case when the threshold p value was 0.50.
Now, accuracy and F1 score have both increased by one percentage point to 91% and
94%, respectively, while precision has increased by two percentage points to 90%.

Exhibit 48  Confusion Matrix of Model Results for CV Data with Threshold p

Value = 0.60
Confusion Matrix for CV Data with Threshold = 0.6

Actual Training Labels

Class “1” Class “0”
Predicted Class “1” 284 (TP) 30 (FP)
Results Class “0” 7 (FN) 114 (TN)

Performance Metrics

TP = 284, FP = 30, FN = 7, TN = 114

P = 284 / (284+30) = 0.90
R = 284 / (284+7) = 0.98
F1 Score = (2 × 0.90 × 0.98) / (0.90 + 0.98) = 0.94
Accuracy = (284 + 114) / (284 + 30 + 114 + 7) = 0.91

7.4  Results and Interpretation

The final ML model with the appropriate threshold p value has been validated and is
now ready for use. The model can be used to predict the sentiment of new sentences
from the test data corpus as well as new sentences from similar financial text data
sources, such as news wires, earnings call transcripts, and quarterly financial reports.
The final model is a collection of penalized regression coefficients for unigram and
bigram tokens from the BOW of the training corpus. To use the model to predict the
sentiment of new sentences, tokenization and identical cleansing and preprocessing
operations must be performed on the new sentences. All the processes performed
on the training data must be performed on the new data to which the model will be
applied (as was done for the test dataset). The model will use the trained penalized
regression coefficients on the term frequency (TF) values of the tokens in the document
 © CFA Institute. For candidate use only. Not for distribution.
Financial Forecasting Project: Classifying and Predicting Sentiment for Stocks 581

without reading large documents. These sentiment classifications can also be used
as structured input data for larger ML models that have a specific purpose, such as
to predict future stock price movements.

EXAMPLE 7 

Comparing Performance Metrics for Confusion Matrixes

with Different Threshold p Values
In the previous analysis using the cross-­validation dataset, performance mea-
sures for the sentiment classification ML model were calculated for a wide
range (from 0.45 to 0.75) of threshold p values. The threshold value of 0.60 was
determined to be the p value that maximizes model accuracy and F1 score; the
confusion matrix for this model is shown in Exhibit 48. Use the following con-
fusion matrixes with threshold p values of 0.75 and 0.45, A and B, respectively,
to answer the following questions.
Confusion Matrix A Confusion Matrix B
Confusion Matrix for CV, Threshold = 0.75 Confusion Matrix for CV, Threshold = 0.45
N = 436 Actual Training Labels N = 436 Actual Training Labels
Class “1” Class “0” Class “1” Class “0”
Predicted Class “1” 281 28 Predicted Class “1” 281 41
Results Class “0” 17 110 Results Class “0” 4 110

Performance Metrics Performance Metrics

TP = 281, FP = 28, FN = 17, TN = 110 TP = 281, FP = 41, FN = 4, TN = 110

Precision = TP/(TP + FP) = 0.91 Precision = TP/(TP + FP) = 0.87
Recall = TP/(TP + FN) = 0.94 Recall = TP/(TP + FN) = 0.99
F1 Score = HMean: Prec. & Recall = 0.93 F1 Score = HMean: Prec. & Recall = 0.93
Accuracy = (TP + TN)/N = 0.90 Accuracy = (TP + TN)/N = 0.90

1 Compare the performance metrics of confusion matrix A (using a thresh-

old p value of 0.75) with the confusion matrix in Exhibit 48 (using a
threshold p value of 0.60).
2 Compare the performance metrics of confusion matrix B (using a thresh-
old p value of 0.45) with the confusion matrix in Exhibit 48 (using a
threshold p value of 0.60).
3 Contrast the performance metrics of confusion matrixes A and B, and
explain the trade-­offs implied between them.

Solution to 1:
Since confusion matrix A has fewer true positives (TPs) and fewer true neg-
atives (TNs) than the confusion matrix in Exhibit 48 (281 vs. 284 and 110 vs.
114, respectively), confusion matrix A has lower accuracy and a lower F1 score
compared to the one in Exhibit 48 (0.90 vs. 0.91 and 0.93 vs. 0.94, respectively).
Also, although confusion matrix A has slightly better precision, 0.91 vs. 0.90,
due to a few less false positives (FPs), it has significantly lower recall, 0.94 vs.
0.98, due to having many more false negatives (FNs), 17 vs. 7, than the confusion
matrix in Exhibit 48. On balance, the ML model using the threshold p value of
0.60 is the superior model for this sentiment classification problem.
Solution to 2:
Confusion matrix B has the same number of TPs (281) and TNs (110) as con-
fusion matrix A. Therefore, confusion matrix B also has lower accuracy (0.90)
and a lower F1 score (0.93) compared to the one in Exhibit 48. Although con-
fusion matrix B has slightly better recall, 0.99 vs. 0.98, due to fewer FNs, it has
somewhat lower precision, 0.87 vs. 0.90, due to having many more FPs, 41 vs.
 © CFA Institute. For candidate use only. Not for distribution.
582 Reading 8 ■ Big Data Projects

30, than the confusion matrix in Exhibit 48. Again, it is apparent that the ML
model using the threshold p value of 0.60 is the better model in this sentiment
classification context.
Solution to 3:
The main differences in performance metrics between confusion matrixes A
and B are in precision and recall. Confusion matrix A has higher precision, at
0.91 vs. 0.87, but confusion matrix B has higher recall, at 0.99 vs. 0.94. These
differences highlight the trade-­off between FP (Type I error) and FN (Type II
error). Precision is useful when the cost of FP is high, such as when an expen-
sive product that is fine mistakenly fails quality inspection and is scrapped; in
this case, FP should be minimized. Recall is useful when the cost of FN is high,
such as when an expensive product is defective but mistakenly passes quality
inspection and is sent to the customer; in this case, FN should be minimized.
In the context of sentiment classification, FP might result in buying a stock for
which sentiment is incorrectly classified as positive when it is actually negative.
Conversely, FN might result in avoiding (or even shorting) a stock for which
the sentiment is incorrectly classified as negative when it is actually positive.
The model behind the confusion matrix in Exhibit 48 strikes a balance in the
trade-­off between precision and recall.

SUMMARY
In this reading, we have discussed the major steps in big data projects involving the
development of machine learning (ML) models—namely, those combining textual big
data with structured inputs.
■■ Big data—defined as data with volume, velocity, variety, and potentially lower
veracity—has tremendous potential for various fintech applications, including
several related to investment management.
■■ The main steps for traditional ML model building are conceptualization of the
problem, data collection, data preparation and wrangling, data exploration, and
model training.
■■ For textual ML model building, the first four steps differ somewhat from those
used in the traditional model: Text problem formulation, text curation, text
preparation and wrangling, and text exploration are typically necessary.
■■ For structured data, data preparation and wrangling entail data cleansing and
data preprocessing. Data cleansing typically involves resolving incompleteness
errors, invalidity errors, inaccuracy errors, inconsistency errors, non-­uniformity
errors, and duplication errors.
■■ Preprocessing for structured data typically involves performing the following
transformations: extraction, aggregation, filtration, selection, and conversion.
■■ Preparation and wrangling text (unstructured) data involves a set of text-­
specific cleansing and preprocessing tasks. Text cleansing typically involves
removing the following: html tags, punctuations, most numbers, and white
spaces.
 © CFA Institute. For candidate use only. Not for distribution.
Practice Problems 585

the financial data using normalization. She notes that over the full sample dataset, the
“Interest Expense” variable ranges from a minimum of 0.2 and a maximum of 12.2,
with a mean of 1.1 and a standard deviation of 0.4.
Steele and Schultz then discuss how to preprocess the raw text data. Steele tells
Schultz that the process can be completed in the following three steps:
Step 1 Cleanse the raw text data.
Step 2 Split the cleansed data into a collection of words for them to be
normalized.
Step 3 Normalize the collection of words from Step 2 and create a distinct set
of tokens from the normalized words.
With respect to Step 1, Steele tells Schultz:
“I believe I should remove all html tags, punctuations, numbers, and extra
white spaces from the data before normalizing them.”
After properly cleansing the raw text data, Steele completes Steps 2 and 3. She
then performs exploratory data analysis. To assist in feature selection, she wants to
create a visualization that shows the most informative words in the dataset based on
their term frequency (TF) values. After creating and analyzing the visualization, Steele
is concerned that some tokens are likely to be noise features for ML model training;
therefore, she wants to remove them.
Steele and Schultz discuss the importance of feature selection and feature engi-
neering in ML model training. Steele tells Schultz:
“Appropriate feature selection is a key factor in minimizing model over-
fitting, whereas feature engineering tends to prevent model underfitting.”
Once satisfied with the final set of features, Steele selects and runs a model on the
training set that classifies the text as having positive sentiment (Class “1” or negative
sentiment (Class “0”). She then evaluates its performance using error analysis. The
resulting confusion matrix is presented in Exhibit 2.

Exhibit 2  Confusion Matrix

Actual Training Results

Class “1” Class “0”
Predicted Class “1” TP = 182 FP = 52
Results Class “0” FN = 31 TN = 96

1 Which of Steele’s statements relating to the steps in building structured data-­

based and text-­based ML models is correct?
A Only Statement 1 is correct.
B Only Statement 2 is correct.
C Statement 1 and Statement 2 are correct.
2 Steele’s concern about using Twitter data in the model best relates to:
A volume.
B velocity.
C veracity.
3 What type of error appears to be present in the IPO Date column of Exhibit 1?
A invalidity error.
 © CFA Institute. For candidate use only. Not for distribution.
586 Reading 8 ■ Big Data Projects

B inconsistency error.
C non-­uniformity error.
4 What type of error is most likely present in the last row of data (ID #4) in
Exhibit 1?
A Inconsistency error
B Incompleteness error
C Non-­uniformity error
5 During the preprocessing of the data in Exhibit 1, what type of data transforma-
tion did Steele perform during the data preprocessing step?
A Extraction
B Conversion
C Aggregation
6 Based on Exhibit 1, for the firm with ID #3, Steele should compute the scaled
value for the “Interest Expense” variable as:
A 0.008.
B 0.083.
C 0.250.
7 Is Steele’s statement regarding Step 1 of the preprocessing of raw text data
correct?
A Yes.
B No, because her suggested treatment of punctuation is incorrect.
C No, because her suggested treatment of extra white spaces is incorrect.
8 Steele’s Step 2 can be best described as:
A tokenization.
B lemmatization.
C standardization.
9 The output created in Steele’s Step 3 can be best described as a:
A bag-­of-­words.
B set of n-­grams.
C document term matrix.
10 Given her objective, the visualization that Steele should create in the explor-
atory data analysis step is a:
A scatter plot.
B word cloud.
C document term matrix.
11 To address her concern in her exploratory data analysis, Steele should focus on
those tokens that have:
A low chi-­square statistics.
B low mutual information (ML) values.
C very low and very high term frequency (TF) values.
12 Is Steele’s statement regarding the relationship between feature selection/fea-
ture engineering and model fit correct?
A Yes.
B No, because she is incorrect with respect to feature selection.
C No, because she is incorrect with respect to feature engineering.
 © CFA Institute. For candidate use only. Not for distribution.
Solutions 589

7 B is correct. Although most punctuations are not necessary for text analysis and
should be removed, some punctuations (e.g., percentage signs, currency sym-
bols, and question marks) may be useful for ML model training. Such punctu-
ations should be substituted with annotations (e.g., /percentSign/, /dollarSign/,
and /questionMark/) to preserve their grammatical meaning in the text. Such
annotations preserve the semantic meaning of important characters in the text
for further text processing and analysis stages.
8 A is correct. Tokenization is the process of splitting a given text into separate
tokens. This step takes place after cleansing the raw text data (removing html
tags, numbers, extra white spaces, etc.). The tokens are then normalized to cre-
ate the bag-­of-­words (BOW).
9 A is correct. After the cleansed text is normalized, a bag-­of-­words is created. A
bag-­of-­words (BOW) is a collection of a distinct set of tokens from all the texts
in a sample dataset.
10 B is correct. Steele wants to create a visualization for Schultz that shows the
most informative words in the dataset based on their term frequency (TF, the
ratio of the number of times a given token occurs in the dataset to the total
number of tokens in the dataset) values. A word cloud is a common visual-
ization when working with text data as it can be made to visualize the most
informative words and their TF values. The most commonly occurring words
in the dataset can be shown by varying font size, and color is used to add more
dimensions, such as frequency and length of words.
11 C is correct. Frequency measures can be used for vocabulary pruning to remove
noise features by filtering the tokens with very high and low TF values across
all the texts. Noise features are both the most frequent and most sparse (or
rare) tokens in the dataset. On one end, noise features can be stop words that
are typically present frequently in all the texts across the dataset. On the other
end, noise features can be sparse terms that are present in only a few text files.
Text classification involves dividing text documents into assigned classes. The
frequent tokens strain the ML model to choose a decision boundary among the
texts as the terms are present across all the texts (an example of underfitting).
The rare tokens mislead the ML model into classifying texts containing the rare
terms into a specific class (an example of overfitting). Thus, identifying and
removing noise features are critical steps for text classification applications.
12 A is correct. A dataset with a small number of features may not carry all the
characteristics that explain relationships between the target variable and the
features. Conversely, a large number of features can complicate the model and
potentially distort patterns in the data due to low degrees of freedom, causing
overfitting. Therefore, appropriate feature selection is a key factor in minimiz-
ing such model overfitting. Feature engineering tends to prevent underfitting in
the training of the model. New features, when engineered properly, can elevate
the underlying data points that better explain the interactions of features. Thus,
feature engineering can be critical to overcome underfitting.
13 A is correct. Precision, the ratio of correctly predicted positive classes (true
positives) to all predicted positive classes, is calculated as:
Precision (P) = TP/(TP + FP) = 182/(182 + 52) = 0.7778 (78%).

14 B is correct. The model’s F1 score, which is the harmonic mean of precision and
recall, is calculated as:
F1 score = (2 × P × R)/(P + R).
F1 score = (2 × 0.7778 × 0.8545)/(0.7778 + 0.8545) = 0.8143 (81%).
 © CFA Institute. For candidate use only. Not for distribution.
590 Reading 8 ■ Big Data Projects

15 A is correct. The model’s accuracy, which is the percentage of correctly pre-

dicted classes out of total predictions, is calculated as:
Accuracy = (TP + TN)/(TP + FP + TN + FN).
Accuracy = (182 + 96)/(182 + 52 + 96 + 31) = 0.7701 (77%).