CHAPTER 1:

● The trend is the long-term movement in the data, showing an overall

increase or decrease over time.
It reflects the general direction in which the data is moving.
Example: A country's GDP generally follows an upward trend over
several decades.
● Seasonality repeating patterns that occur at regular intervals (daily,
monthly, yearly).
It is predictable because it follows a fixed calendar period.
Example: Ice cream sales increase in summer and drop in winter every
year.
● Unlike seasonality, cyclical variations do not follow a fixed calendar
period but occur over time, often influenced by economic or business
cycles.
● Noise is the unpredictable variation in data that doesn’t follow a
pattern.

It is caused by random shocks, external disturbances, or

measurement errors.

It is short-lived and does not repeat over time.

Example: A sudden economic crisis or a natural disaster impacting

stock prices.

● A trend with a seasonal pattern means that the data exhibits both a
long-term increase or decrease (trend) and short-term, repeating
fluctuations at fixed intervals (seasonality). This means:
1. The trend shows a general upward or downward movement over time.
2. The seasonality introduces periodic variations that repeat at regular
intervals (e.g., daily, monthly, yearly).

● stationary time series does not have a trend or seasonality—its

statistical behavior remains the same throughout (mean and variance are
constant over time) and fluctuates around a constant mean.
Expert Judgment and Delphi Method are both qualitative forecasting
techniques that rely on expert opinions, but they differ in structure and
process:

● Expert Judgment:

○ Involves obtaining forecasts from one or a few experts.

○ Often unstructured and relies on the individual’s personal
experience and intuition.
○ May be faster to gather but can be more subjective and prone to
individual biases.

Expert judgment leverages an individual expert’s unique experience,

intuition, and domain-specific knowledge to make forecasts.

● Delphi Method:

○ Involves a panel of experts who provide forecasts independently.

○ Uses a structured, iterative process where experts share their
opinions anonymously.
○ After each round, a summary of the forecasts is provided,
allowing experts to reconsider and refine their opinions.
○ Aims to reduce bias and converge towards a consensus forecast.

In summary, while expert judgment depends on individual insight, the Delphi

method systematically aggregates multiple expert opinions to mitigate
individual biases and improve forecast reliability.

CHAPTER 2:
● Autocorrelation (ACF) and partial autocorrelation (PACF) are
statistical measures that help analyze the relationship between time
series and its lagged values

==Autocorrelation and partial autocorrelation are two fundamental

concepts in time series analysis that help us understand how current
values relate to past values.
● time series observation = current observation
● Covariance and autocorrelation are both measures of how variables
relate to each other, but they differ in focus and normalization:

Covariance

Definition:
Covariance measures how two random variables vary together.

Scope:
It applies to any two variables (they can be different).

Units:
Its value is not standardized and depends on the units of the variables.

Interpretation:
A positive covariance indicates that the variables tend to move in the
same direction, while a negative covariance indicates they move in
opposite directions.

Autocorrelation

Definition:
Autocorrelation is a specific application of covariance to time series data.
It measures the correlation of a time series with a lagged version of itself.

Scope:
It focuses on the relationship within the same variable over time,
capturing patterns like seasonality or persistence.

Interpretation:
It tells you how much past values of a series are linearly related to its
current values.

● Yt: current values of the time series

● Yt−k: past values of the time series
● k is the lag (how many time steps back we are looking)
● autocorrelation = covariance(yt,yt-k)/ variance(yt)
● variance(yt): covariance at lag 0

● A decaying pattern generally means that the influence or effect of

a past event or observation diminishes over time
When analyzing a time series, a decaying autocorrelation pattern
means that the correlation between the current value and its past
values decreases as the time lag increases.
● Partial Autocorrelation Function (PACF): The PACF measures the
correlation between Yt and Yt-k after removing the effects of intervening
observations Yt-1,Yt-2,....Yt-(k-1)
● For k≥2 ; Φkk is defined as the correlation between the prediction errors
of Yt and Yt−kafter removing the linear effects of the intervening
variables.
● ρ1is the autocorrelation at lag 1

● ACF tells you the overall correlation between past and current values,
including both direct and indirect effects.

● PACF focuses solely on the direct impact of a specific lag by removing

the influence of the intervening data, using the prediction errors from
prior regressions. (PACF looks at these mistakes to figure out the direct
relationship.)
● PACF then correlates this residual with the temperature k days ago (e.g.,
two days ago). This correlation tells you the direct effect of that past day
on today’s temperature, independent of the intervening days.

● Imagine you're trying to predict today's temperature using temperatures

from previous days. When you make this prediction, there will be some
error— the difference between what you predicted and the actual
temperature.
The error (or residual) from this prediction represents what the earlier
days couldn’t explain.
● When you predict today's temperature using multiple past days, some
days influence the prediction directly while others might only contribute
indirectly through their relationship with other days.

A stochastic process:
● is composed of random variables, typically denoted as X(t) where t
represents time. Each X(t) captures the state of the process at time t.

The process evolves according to a set of probabilistic laws. This

means that the way the process changes from one time point to the
next is governed by probability rather than a fixed, deterministic rule.
● Probabilistic Laws:

The evolution of the process is dictated by probabilities.

white noise process:

● is a fundamental type of stochastic process used in time series
analysis. It consists of a sequence of independent and identically
distributed (i.i.d.) random variables, meaning that each value in the
sequence is generated randomly, without any dependency on past
values.
● Independence: Each observation in the process does not depend on
previous observations.
● Identically Distributed: Each value in the sequence follows the same
probability distribution, often assumed to be a normal distribution with
mean 0 and variance σ^2. (process has 0 mean)

expected value =mean

This means the process has a constant mean and variance over time.

● Example of a White Noise Process:

If we simulate a white noise process, it might look like this:

Xt={−0.5,1.2,−0.3,0.8,−1.1,0.4,−0.7,1.5,−0.2,0.6,…}

Each value appears random, fluctuates around zero, and has no

predictable pattern.

● The autocorrelation function (ACF) of a stochastic process measures

the correlation between the values of the process at different time
points.
● When we say "a value is always perfectly correlated with itself", we
mean that any random variable is fully predictable when compared to
itself.

Correlation measures the strength of the relationship between two

variables.

When k=0 (i.e., comparing a value with itself), the numerator is simply

Var(εt)=σ^2, so: ρ(0)=σ2/σ2=1.

any variable is always fully correlated with itself. ρ(0)=1

● No correlation between different time points ρ(k)=0 for k≠0

 For a white noise process, the values at different time points are
independent : no linear relationship

● The noise term εt is usually assumed to be part of a white noise

process.

Random walk process:

● the random noise terms in a random walk are not derived from the past
values of the series but are introduced as new, unpredictable shocks at
each time step, typically modeled as white noise.
● These shocks can represent various types of disturbances like
economic shocks, environmental factors, or any other unpredictable
event that affects the system.

● The random walk process {Yt} is defined as:

Y1=e1
Y2=e1+e2
Yn=e1+e2+...+en
Or, more generally:
Yt=Y(t−1)+et , for t>1 Yt is the sum of all past shocks

● the lag k represents how far apart two observations Yt and

● When 𝑘 is small (e.g., 1, 2, or 3), 𝑌𝑡 and yt-k are close together in time,
Yt−k are in time.

meaning their values tend to be similar.

This results in high autocorrelation for small values of k.

● As k increases, we compare observations that are further apart in

time: less correlated and tend to behave more like independent
values, so ρ(k) approaches 0.
● the assumption that when k is close to 0, ρ(k) is close to 1 applies in
cases where the time series exhibits persistence or slow decay of
correlation over time.

● This explanation refers to the behavior of the autocorrelation function

(ACF) in a stationary time series where observations that are closer in
time tend to be more similar (high correlation), while those farther apart
tend to be less related (low correlation). ??????
● In a stationary time series, the autocorrelation function (ACF) typically
shows a gradual decay as the lag k increases. ( k neb3ed aal haja l
effect mteeha yetnaha)

This means that observations that are close together in time tend to
be more similar (high correlation), while observations that are farther
apart in time tend to be less related (low correlation).

However, the exact rate at which the correlation decays depends on

the specific time series process.

● random walk is used to model stock prices , animal locations ,

movements of molecules in gases and liquids

Moving average process:

● a time series model where the value at time t is determined by the
weighted sum (average) of past random shocks (white noise)(error).

Autoregressive process:
● the present value depends on the previous value but with a
reduced impact. (of the error term)
● ro : determines how much of the previous value influences Yt
● if ro (closer to 1): Yt would be strongly dependent on Yt−1,
leading to a persistent trend.
● if ro (closer to 0), the process would behave more like white
noise.

Stationarity:
● Statistical properties of a time series do not change over time
● does not have a trend or seasonality—its statistical behavior remains
the same throughout (mean and variance are constant over time) and
fluctuates around a constant mean.
● The autocovariance depends only on the time lag (time
difference) not on the specific time 𝑡.
Yes! A time series with no trend and no seasonality is a strong candidate
for being stationary. However, to be strictly stationary, it must satisfy
the following conditions:

1. Constant Mean – The average value of the series does not change
over time.
2. Constant Variance – The fluctuations (spread of values) remain
consistent over time.
3. Constant Autocovariance – The relationship between
observations at different time points (lag structure) remains the
same over time.

>>> stochastic process Yt is weakly stationary

● weakly stationary =second-order stationary
> definition of stationarity that focuses on the first two
moments of a stochastic process — mean and variance, and
the autocovariance.
> called "weak" because it does not require all moments or
higher-order statistics to be constant over time, only these
basic properties.

● White noise process, autoreggresive process and moving average

process are stationary
● Random walk process is not stationary

Why is a Random Walk Not Stationary?

 ● In a random walk process, the mean increases over time because
Yt is a cumulative sum of past noise terms.
● As time progresses, the random walk "drifts" away from its
starting point due to the cumulative effect of random shocks. This
means the mean of the process is not constant over time

CHAPTER 3:
● Exponential smoothing reduces the weight of older observations by
applying an exponentially decreasing weight factor. This is done using a
smoothing parameter, α\alpha (0 < α ≤ 1), which determines how
much weight is given to the most recent observation versus past
observations.
● As k increases, (1−α)^k gets smaller, meaning older
observations contribute less to the forecast.

● The smoothed value St is the estimate of the underlying pattern in the

data at time t. It serves as the forecast for the next time step in Simple
Exponential Smoothing (SES).

St is not just the actual observed value Xt; instead, it is an adjusted

version of past observations.

● Xt is raw data, while St is a smoothed estimate

* Xt represents real-world observations, which may contain noise or

sudden jumps.

* St filters out noise by averaging past values using exponential

smoothing.

● St reacts more slowly to changes than Xt

 * If the data has random fluctuations, St will smooth them out.

* If Xt suddenly spikes, St will adjust gradually instead of jumping

immediately.

● Random fluctuations, also called noise, refer to unpredictable

variations in a time series that do not follow any pattern, trend, or
seasonality.
● initial smoothed value = initial actual value

● Why Does a Smaller α Apply More Smoothing?

alpha ontrols how much weight is given to the most recent observation
in SES.

● A larger α gives more weight to the latest observation Xt,

making the forecast react quickly to changes.
● means the model focuses mostly on the latest data, making it
more responsive to recent changes.
● α=0.9 → A jagged line, closely following the actual data.

● A smaller α gives less weight to Xt and more weight to past

values, making the forecast smoother and slower to react to
sudden changes.
● means the model remembers past values more and smooths out
fluctuations.
● α=0.1 → A smooth curve, reducing random fluctuations.

● To smooth out fluctuations means to reduce or minimize the impact of

sudden changes, spikes, or dips in data or trends, making them appear
more stable or gradual.
● The choice of the smoothing parameter (alpha, α) indeed impacts the
sensitivity of the model. (sensitivity to recent changes and
fluctuations)
● h: A forecast horizon is the future time period over which predictions
or forecasts are made. It defines how far into the future you want to
forecast.
● The level (baseline value) in time series analysis represents the
average or typical value of the data series at a given point in time,
excluding any trends or seasonal effects.
● Double Exponential Smoothing (DES) extends Simple Exponential
Smoothing (SES) by adding a component to capture trends in time
series data. It uses two smoothing parameters to handle both the
level (baseline value) and the trend (rate of change) of the data.

● Trend smoothing factor: β

● Smoothing factor: α

● Triple Exponential Smoothing:

Can be additive or multiplicative.

🔄 Choosing Between Additive and Multiplicative:

● Additive: Seasonal patterns are constant over time (fixe

increase/decrease).

example: Sales increase by a fixed amount every winter (e.g., +50

units every December).

● Multiplicative: Seasonal patterns grow or shrink with the level

(percentage-based). (change over time)

example: Sales double every winter compared to the average — a fixed

percentage change.

● Use ETS models when you need a more flexible approach

that captures interactions between error, trend, and
seasonality.
CHAPTER 4:

● A random shock is an unexpected fluctuation or disturbance in the

data at a specific time point.
● In time series terms, these shocks are often referred to as residuals (ϵt
) and are assumed to be white noise:

Mean: 0.

Constant variance: σ2\sigma^2σ2.

No autocorrelation: Errors are independent over time

● A Moving Average (MA) process in time series analysis depends

on past random shocks (errors) rather than past values of the
series itself because it is designed to model the impact of
unpredictable, short-term fluctuations directly on the current value
of the series
NOTES:

● A cyclical pattern in time series:Is not necessarily linked to specific

seasonal effects.
● advanced models (like ARIMA) are typically used for cyclical patterns.
● seasonal patterns:Repeats at fixed intervals every year
● Time series forecasting is used to:

○ Identify trends in historical data

○ Predict future values based on past observations

○ Make strategic decisions

● Quantitative forecasting models include:

○ Causal models

○ Time series models

● A time series forecasting model should be chosen based on:

○ Type of time series data

○ Complexity of the data

○ Desired level of accuracy

● Which of the following statements about ARIMA models are true?

○ They are used for time series forecasting

○ They capture trends and seasonality

○ They use a combination of autoregression and moving averages

● The autocorrelation function (ACF) is used to:

○ Measure the strength of relationships between time series

observations at different lags
 ○ Determine if a time series is stationary (ACF should drop rapidly
(often to zero) after a few lags)

○ Identify significant lags for forecasting models(The ACF plot shows

spikes at different lags)

● A weakly stationary time series has:

● Constant mean over time (0)

● Constant variance over time

● Autocovariance that depends only on lag k (cst)

● The Augmented Dickey-Fuller (ADF) test is used to check: (pvalue:non)

● Whether a time series is stationary

● If a time series has a unit root

● The KPSS test:

● Tests whether a series is trend-stationary

● Uses p-values to determine stationarity

● A white noise process is characterized by:

● Independent and identically distributed (iid) random variables

● Constant variance over time

● No correlation between different time points

● random walk process:

● Is a time series where each value is derived from the previous value
plus random noise

● Has autocorrelation close to 1 for nearby observations

● Is commonly used to model stock prices

● Simple exponential smoothing (SES) is appropriate when:

● Data has no trend or seasonality

● The smoothing parameter (α) determines the weight of recent

observations
 ● The goal is to track short-term fluctuations

● Double exponential smoothing (Holt’s method):

● Extends SES by incorporating a trend component

● Uses two smoothing parameters (α and β)

● Models linear trends

● If a time series has a trend but no seasonality, the best ETS model is:
ETS(A, A, N)
● The function acf() in R is used for: Calculating autocorrelation
● The main advantage of using deep learning models for time series
forecasting is:

● Ability to capture complex patterns

● Handling long-range dependencies

● Which forecasting methods can be used for stationary time series?

● ARIMA

● Exponential Smoothing

● Moving Average Models