Algorithmic Trading

By: Avi Thaker

What is Algorithmic Trading

The use of electronic platforms

for entering trading orders with
an algorithm which executes
pre-programmed trading
instructions whose variables may
include timing, price, or
quantity of the order.

http://neverlosstrading.com/Algorithmic%20Trading.html
Motivation

Computers can process larger amounts of data than humans and make
decisions faster than humans
Algorithms do what they are told, takes the human emotion out of trading
Trillions of $$$ traded daily - highly paid employees
Bleeding edge of sciences; math, engineering, computer science, etc.
Introduction:
The Efficient Market Hypothesis
Impossible to beat the market
Market efficiency causes prices to incorporate and reflect all relevant information
Prices adjust quickly to new information
Prices should reflect all available information

Securities always trade at their fair value

The only way to obtain higher returns is to purchase riskier investments
Cannot predict trends, history does not indicate future
Warren Buffett? Citadel? Two Sigma?
Are they lucky
The market is becoming increasingly more efficient
Index

DataCollection.jar
Python Pandas
Hello World Algo Trading
Kinds of Trading Strategies
Backtesting
System Architecture
Appendix
DataCollection.jar
Collects intra-day data from Google Finance
Writes data to current directory
Snapshots
Can collect multiple stocks
Type in ticker, hit collect data
Works on Macs and PCs with Java
Example Output (tab delimited)
127.35 4:00PM EDT

Run me please? See how I work.

A programming introduction
Python Pandas
Developed for data manipulation and
analysis by Wes Mckinney at AQR
Extremely fast vector driven manipulation
Dataframe manipulation, integrated indexing
Data alignment and integrated handling of
missing data
Time series-functionality
Date range generation
Frequency conversion
Regressions, moving window statistics
And so much more!
 Pandas Example: Getting Historical Data
#pandas_example.py
# Imports
import datetime
import pandas as pd
import pandas.io.data
I am interactive
from pandas import Series, DataFrame (zoom into me)
import matplotlib.pyplot as plt
import matplotlib as mpl
mpl.rc('figure', figsize=(8, 7))

# Get Data
aapl = pd.io.data.get_data_yahoo('AAPL',
start=datetime.datetime(2006, 10, 1),
end=datetime.datetime(2015, 4, 7) )
aapl.head()
# Do some time series manipulation
aapl['SMA50'] = pd.rolling_mean(aapl['Adj Close'],50)
aapl['EMA50'] = pd.ewma(aapl['Adj Close'], 50)

# Plot
plt.figure()
plot(aapl.index, aapl['Adj Close'])
plot(aapl.index, aapl['SMA50'])
plot(aapl.index, aapl['EMA50'])
plt.legend(('ADJ Close', 'SMA50', 'EMA50'))
 Pandas Example: DataFrame

Data Arranged In columns indexed

by date
Resembles Excel/R
Handles NaN data extremely well # view the first few elements aapl.head()

Great for data manipulation Open High Low Close Volume Adj Close SMA50 EMA50
Date
Common functions 10/2/2006 75.1 75.87 74.3 74.86 178159800 10.09NaN 10.09
10/3/2006 74.45 74.95 73.19 74.08 197677200 9.98 NaN 10.03446
Df.head()
10/4/2006 74.1 75.46 73.16 75.38 207270700 10.16 NaN 10.07714
Df.tail() 10/5/2006 74.53 76.16 74.13 74.83 170970800 10.08NaN 10.07787
10/6/2006 74.42 75.04 73.81 74.22 116739700 10NaN 10.06168
Df.index()
Df.ffill()
Align and fill missing data
Pandas Example: Data Visualization

df = pd.io.data.get_data_yahoo(symbols=['AAPL', 'GE',
'GOOG', 'IBM', 'KO', 'MSFT', 'PEP'])['Adj Close']
rets = df.pct_change()
a = plt.figure()
plt.scatter(rets.PEP, rets.KO)
plt.xlabel('Returns PEP')
plt.ylabel('Returns KO')
pd.scatter_matrix(rets, diagonal='kde', figsize=(10, 10));
Pandas Example: Data Visualization 2

pd.scatter_matrix(rets, diagonal='kde', figsize=(10, 10));

Pandas Example: Data Visualization 3

# Display a Correlation of returns

corr = rets.corr()
plt.figure()
plt.imshow(corr, cmap='hot',
interpolation='none')
plt.colorbar()
plt.xticks(range(len(corr)),
corr.columns)
plt.yticks(range(len(corr)),
corr.columns)
Pandas Example: Data Visualization 4

plt.scatter(rets.mean(), rets.std())
plt.xlabel('Expected returns')
plt.ylabel('Risk')
for label, x, y in zip(rets.columns,
rets.mean(), rets.std()):
plt.annotate(label, xy = (x, y), xytext =
(20, -20),textcoords = 'offset points', ha =
'right', va = 'bottom', bbox = dict(boxstyle =
'round,pad=0.5', fc = 'yellow', alpha = 0.5),
arrowprops = dict(arrowstyle = '->',
connectionstyle = 'arc3,rad=0'))

http://nbviewer.ipython.org/github/twiecki/financial-analysis-
python-tutorial/blob/master/1.%20Pandas%20Basics.ipynb
Hello World Algorithmic Trading
Backtesting a Moving Average Crossover Strategy
http://www.quantstart.com/articles/Backtesting-a-Moving-Average-Crossover-in-
Python-with-pandas, Zipline
Strategy: Moving Average Crossover

Momentum Strategy (more later)

Enter long, when longer MA crosses over
the shorter faster MA
Enter short (or exit), when shorter faster
MA crossover longer MA
Believe that market will strongly move in
one direction
Prone to whiplash
We will build and backtest this!
Getting Started
Prerequisites
Python - 2.7+
NumPy
Pandas
Matplotlib
I suggest using pythonXY/Spyder does everything for you
We will now create this strategy and run it on AAPL
Take ma_cross.py and backtest.py and place them into the same directory
Run the files

Change the Parameters to Short=20, Long=200,

on ARMH from 2011-Now
Code Details and Questions
Only trades long, creates two simple moving averages
MovingAverageCrossStrategy is subclassed from Strategy
MarketOnClosePortfolio is subclassed from Portfolio
What is misleading with the data we are using?
Hint look at AAPLs price.
How can we improve on this code?
How do we measure strategy performance?
How can we tell if a strategy is good?
What are some other kinds of trading strategies?
Output
Python: Quantopian

Extremely Slow, but provides a decent

backtester
Direct Python interface to IB
Live and Paper Trading
Hosts competitions, if you win, will be
awarded with $100,000 to trade with
All profits go to you
https://www.quantopian.com/open
Open source community
https://www.quantopian.com/home
https://www.quantopian.com/posts
 A sample momentum algorithm from
Quantopian.
Very simple coding interface to use, and push
to live trading with IB, or paper trade
This system can directly trade with your
Interactive Brokers account
Must give privileges, have cash, etc.
Sample results from a paper trade, with statistics and an equity curve
Kinds of Strategies
A Brief Introduction
Terms
Depth of Market (DOM, Limit Order Book, Level 2)
Best Bid price to instantly sell
Best Ask price to instantly buy
Bid Size
Ask Size
Alpha
The abnormal rate of return in excess of what would be
predicted by an equilibrium model like the capital
asset pricing model (CAPM).
Tick Size Interval between possible prices
Slippage Difference between signaled price and
trade price
HFT vs Non HFT
HFT or High Frequency Trading, relies on quick
algorithmic trading to capture small spreads or small
alpha (less than 1 minute holds)
Difficult for the retail investor to get into
Extremely competitive
Fastest usually takes all profits
Makes up 70%+ of all trading volume

Non HFT (trades of longer periods)

Easier for the retail investor to get into
Not as competitive
Smarter people make profits
This is our focus
Strategy: Alpha Taking
http://marcoagd.u
suarios.rdc.puc-
rio.br/revers.html
Believes the market will
move in one direction
Mean Reverting price
reverts back to a average
price, Bollinger Bands
Stochastic Math
Trend following / Momentum Stockcharts.com
price will move in one
direction
We tested/built one of these
Impulse response filters
Strategy: Market Making

Capture the Spread Bids Price Asks

Increase Liquidity 100.03 2,1
On both sides of the 100.02 3,7,8
market 100.01 5,2,15

Risky during times of 100 1,2,5

volatility 1,2 99.99
2,5,8 99.98
Must be fast and have
good queue position 3,8,1,5,3 99.97
2,3 99.85
Math is generally more
complicated Lots of machine learning, think Bayesian
networks, and neural networks
The focus is on market microstructure
Strategy: Pair Trading

Assumes two contracts move

together
Revert moves when one moves
against the other
Can be highly risky
2007 Crash?
Spreads blow up
Commonly used by Hedgefunds
High usage of Kalman Filters,
CADF (Co-integrated Augmented
Dickey Fuller) tests,
autocorrelation, etc. http://www.investopedia.com/university/
guide-pairs-trading/pairs-trade-
example.asp
Strategy Detail: Common Signals
Simple Moving Average: SMA
Mean of previous n periods
Smooths data, easier to identify
trends
Common signal

+ 1 + + 1
=


= +

=
 Strategy Detail: Common Signals
Exponential Moving Average: EMA

Infinite impulse response filter

Less lag than SMA
Commonly used signal
2
=
+1
1 + 1 2 + 1 2 +
=
1+ 1 + 1 2+
= +
Trading The EMA

Enter Long: Close > EMA & Prev_Close > EMA_Prev

Enter Short: Close < EMA & Prev_Close < EMA_Prev
 Strategy Detail: Common Signals
Bollinger Bands
Volatility Bands
Based upon standard deviation
Identifies points of reversion
Middle Band = 50-Day SMA
Upper Band = 50-Day SMA + 50-Day SD of Price
Lower Band = 50-Day SMA - 50-Day SD of Price
Example: Trading the Bollinger Bands
More Examples / Technical Indicators

RSI Relative Strength Index

Technical indicators are mostly

Parabolic SAR Parabolic Stop and

useless on their own

Reverse
Price Channels
Must identify something that VWAP Volume Weighted Average Price
happens in the market, and use ZigZag
the indicators (or come up with MACD Moving Average Convergence
your own) to represent that Divergence

something PPO Percentage Price Oscillator

KST - Know Sure Thing
Data visualization is crucial Ultimate Oscillator

Simplicity is usually better Vortex Indicator

The list goes on forever
Backtesting
Really again? How much is there to backtesting?
Backtesting a Strategy
Backtesting with Limit order Execution
Simulate by placing limit orders
Need to check for fills
Complex and requires time
Does not perfectly model slippage
Backtesting with Close execution
Orders filled on close of bar
Subject to bid/ask bounce
Must subtract slippage numbers
More than 2 ticks?
Event Driven
 Limit Order Book
Bid Size Ask Size

Volume By Price

Best Ask

Best Bid

http://www.trade2win.com/articles/
1750-understanding-liquidity-market-
pullbacks
Limit Order Execution

For an order to be executed, a trade must cross below your buy, or a trade happens at your price,
when you have been filled in the queue
Limit Order Execution Place Order
Bids Price Asks
100.03 2,1
100.02 3,7,8
100.01 5,2,15
100 1,2,5
1,2 99.99
2,5,8 99.98
3,8,1,5,3 99.97
2,3 99.85

Place limit
order of 2 lots
at 99.99
Limit Order Execution Book Movement
Bids Price Asks
100.03 2,1
100.02 3,7,8
100.01 5,2,15
100 1,2,5
1,2,5 99.99
2,5,8 99.98
3,8,1,5,3 99.97
2,3 99.85

Fill at 99.99, this Another order is People cancel

becomes removed, placed behind their orders
and position you
advances. A trade
happens
Limit Order Execution Order Fill
Bids Price Asks
100.03 2,1
100.02 3,7,8
100.01 5,2,15
100 1,2,5
2,5 99.99
2 99.98
3,8,1,5,3 99.97
2,3 99.85
After an order is filled you move up in the queue,
until you either are filled or cancel the order
We are now first in the queue
Backtesting a Strategy/Risk
Provide evidence of profitability
Curve fitting/ optimization bias
In-sample vs out-of-sample
Forward looking bias
Risk tolerance

Key Statistics
Average wins :: 0.637 USD
Average loss :: -0.438 USD
# Wins :: 214
# Losses :: 210
# Neutrals :: 3
Win Rate :: 0.501
PPC :: 0.104USD
# Traded :: 427.0
Ann. Sharpe :: 2.335
Backtesting a Strategy
Does the strategy work across many assets?
How many years does it work for? Overall: 1980 2016
Sharpe: 2.12
Does it escape the bid-ask bounce? PPC: 0.13
Wins: 12634
Losses: 10527
Risk Tolerance? Trades: 23666

Maximum Drawdown?
Fees? Trading frequency?

Would you trade me?

Sharpe:1.299
PPC:0.338
Wins:255
Losses:202 In Sample: SPY 2004-2010
Trades:463.0 Out of Sample: Assets Randomly Selected:
ADBE XLNX BBBY CFN EMC ADP AFL DE T SPLS DG ADS
ALL MET CL PX WYN
Order Sizing

Not Volatility Sized Orders Volatility Sized Orders

Overall: 2001 2016 Overall: 2001 2016

Sharpe: 2.38 Sharpe: 2.91
PPC: 0.19 PPC: 0.12
Wins: 23448 Wins: 23448
Losses: 19719 Losses: 19719
Trades: 43378 Trades: 43378

Generally size orders inversely proportional to volatility

Biases and Pitfalls
These can be done unintentionally
Curve Fitting Bias
Adjusting/adding parameters until the strategy looks attractive in backtest
Forward looking bias
Program looks at future due to bug in code
Calculating optimal parameters, optimizations
Looking at the data!
Survivorship Bias
Not including full universe (pre 2008 crash, 2007 algo trading blow up)
Psychological Bias
Can you tolerate a 5 month drawdown? Lose half your portfolio
Your backtests will suggest possible severity
System Architecture
How do I actually build an autonomous trading system?
System Architecture Overview

Client (IB) Strategy Backtester

Multi-Threaded
C++, Java, Python Can instantiate multiple
Communication strategies
over interfaces RSI
Event Driven Backtester
Eliminates errors
Can use the same strategy
for trading and
backtesting
Visualizing an Intraday Strategy

Exit Long

Entry Long
General Tips
This is not a get rich quick scheme
Finding alpha is hard, do not get discouraged
Drawdown are painful, be careful with leverage
Trust your alpha (if you have some), strategies are usually simple
Performance
Out of sample performance is generally of in sample performance
Live trading performance is generally of in sample performance
Due to curve fitting, unexpected slippage, etc.
Make sure you account for transaction fees and slippage and order sizes
Fun and exciting way to learn not only the markets but also computer science
and math
Data is your friend
Build your own backtester/execution environment
Questions?
Further Readings
Best guide to starting algo trading (intro/backtester taken from here)
http://www.quantstart.com/
Execution Environment/Backtester/Community
https://www.quantopian.com/
Cheap trading platform with API
https://www.interactivebrokers.com/ind/en/main.php
Stellar documentation on how to do execution

Technical Analysis Library TA-Lib

http://ta-lib.org/
https://pypi.python.org/pypi/TA-Lib
Data:
Free: Yahoo Finance, Google Finance error prone
Cheap: Pi Trading, Kibot, Tickwrite
Appendix
Appendix: Candlestick/Bar Data

Open price at start of bar

High highest price
Low lowest price
Close price at end of bar
Volume number traded during bar
Can be on any timescale: seconds
to monthly

http://www.financial-spread-
betting.com/course/candle-stick-charting.html
Appendix: Order Sizing

Average True Range Scaling

Reduces trade size during times of volatility, Increase during low volatility
Increases Sharpe Ratio
Can adjust to size of contract, and/or contract price

= $1,000

=
10 ($)
= max , ,
1 1 +
=

Appendix: Sharpe Ratio

=

=
=
=
Developed by William F. Sharpe
Measures risk adjusted performance
Risk vs. Reward
Higher is usually better
Usually annualized
Adjust portfolio returns to a daily basis, average the returns and multiply
by 252
Appendix: PPC
Profit Per Contract


=
=
=
A measure of profitability, measured in ticks
A highly liquid stock usually has a tick size of a penny
If your strategy has more than 2 ticks, it is considered profitable (can
escape the bid/ask bounce), if testing on bar data without limit order
execution on bar closes
You can submit market orders and still make money
Assumes liquidity!!!!!
Appendix: CAPM
Capital Asset Pricing Model
= +
=
=
=
=
Describes the relationship between risk and the expected
return
Investors need to be compensated for time (risk free rate)
and risk (beta)
Appendix: Drawdown

The measure of the largest drop from peak to bottom (in percentage)
It is a pain index measure
Extremely important to measure the duration of the drawdown
Do you want to be losing money for years?
= max {() }
(0,)

MDD = max(0,) [max(0,) { ()}]

Where = , 0 is a random process
Simply put maximum drawdown is:
(Peak value before largest drop lowest value before new high) / Peak value before drop
 Appendix: Underwater Curve

Good way to visualize

how much of the time
you are in a drawdown
Lets you evaluate how
much pain you should
be able to handle

http://ctaperformance.com/wntn
Appendix: Distribution of Returns

http://ctaperformance.com/wntn
Generally a histogram of returns
Look at center, shape, distribution, spread
Want positive center, and no major outliers
Appendix: Strategy Correlation

Generally you want to make sure that your strategies are not correlated to
each other (look at daily returns)
You do not want everything to have a bad day at the same time
Balanced returns are good
Uncorrelated strategies tend to yield higher Sharpe ratios when mixed
Correlated strategies tend to reflect the same alpha
These strategies tend to compete with each other
Negatively correlated strategies can be good
Highly negatively correlated strategies can indicate problems with your alpha

Thank you Aaron Rosen for your feedback

Appendix: Tradable AUM

Not all strategies are created equal

Strategy A might be able to trade $1,000,000 without incurring large slippage
but trading $100,000,000 it might incur much more slippage and kill the
strategy
Market making your ability to capture the inside bid offer decreases with size
High frequency strategies
Some momentum strategies
Sharpe ratios and AUM tradable are usually inversely correlated
There are some exceptions
Note that these numbers are artificial

Thank you Aaron Rosen for your feedback

Appendix: Machine Learning

Neural Networks (with and

without reinforcement), support
vector machines, clustering,
Bayesian networks
Garbage in == Garbage out
You cannot expect this tool to
give you good results from
nothing
You are data mining!
Be careful of over-fitting data
Be sure to test, in-sample, out-
of-sample and out-out-of-sample
Appendix: RSI
Relative Strength Index
Momentum indicator
Compares recent gains to recent
losses to determine oversold and
overbought conditions
100
= 100
1 +

=