MScFE xxx [Course Name] - Module X: Collaborative Review Task

Revised: 09/07/2019
© 2019 - WorldQuant University – All rights reserved.
1
MScFE 650 Machine Learning in Finance– Table of Contents

Module 1: Introduction to Machine Learning ............................................... 3

Unit 1: Introduction to Machine Learning ...................................................................... 4

Unit 2: The Community, Projects, Events and Resources ....................................... 14

Unit 3: Machine Learning in Financial Contexts ...................................................... 22

Unit 4: Academic Papers in Review ............................................................................... 28

Bibliography ............................................................................................................................ 34

Collaborative Review Task ................................................................................................ 36

© 2019 - WorldQuant University – All rights reserved.

2
MScFE 650 Machine Learning in Finance - Module 1: Summary

Module 1: Introduction to Machine Learning

Beginning with an overview of the course as a whole, the first module invites you to explore the
field of machine learning as it applies in technology and financial companies. As a general
introduction, the module seeks to describe how machine learning is used, while outlining key
concepts and categories that fall within the field, and familiarizing you with the wider machine
learning community.

© 2019 - WorldQuant University – All rights reserved.

3
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

Unit 1: Introduction to Machine Learning

Definitions of machine learning

There are two common definitions of machine learning:

1 The older definition, described by Arthur Samuel: “The field of study that gives computers the
ability to learn without being explicitly programmed.”
2 A more modern definition by Tom Mitchell: “A computer program is said to learn from
experience E with respect to some class of tasks T and performance measure P, if its
performance at tasks in T, as measured by P, improves with experience E.”

The following is an example of the second definition:

E = Playing checkers

T = Percent of games won against opponents

P = Playing practice games against itself

It is the science of getting computers to act without being explicitly programmed to do so.

History of machine learning

© 2019 - WorldQuant University – All rights reserved.

4
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

© 2019 - WorldQuant University – All rights reserved.

5
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

© 2019 - WorldQuant University – All rights reserved.

6
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

© 2019 - WorldQuant University – All rights reserved.

7
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

Machine learning algorithms

There are three main classes of machine learning algorithms:

1 Supervised learning

The first class we will look at is supervised learning, which refers to machine learning that is based
on a training set of labeled examples.

A supervised learning model trains on a dataset containing features that explain a target variable.
“The object of this kind of learning is for the system to extrapolate, or generalize, its responses so
that it acts correctly in situations not present in the training set.” – Sutton & Barto (1998).

There are two different types of tasks in general, namely classification and regression tasks.

1.1 Classification Tasks

Firstly, the goal of classification tasks is to learn a mapping from inputs x to outputs y, where y ∈
{1, . . . , C}. Here C denotes the number of classes we are using for classification, for example C = 2.
This is called binary classification, in which case we often assume y ∈ {0,1}. Classification tasks,
where C > 2, are called multi-class classification. Another form that classification tasks can take is
multi-label classification, which means that the classes we are using are not mutually exclusive.
Multi-label classification “is best viewed as predicting multiple related binary class labels (a so-
called multiple output model).” – Murphy (2012)

The vast majority of papers about applying supervised learning to finance involve classification-
type problems. For example, in finance this may be to classify stock according to their returns
which can either move up or down. In this context the system is not predicting the price or returns,
but instead classifying whether it moves above or below a certain threshold (e.g. 5% increase or
decrease). Accordingly, the data would then be labeled as either 1 (an up-day), –1 (a down-day), or
0 (no condition met).

Another example is the use of propensity models for targeted marketing. For instance, retail banks
or insurance companies may use customer data (e.g. age, bureau rating, education, and salary) as
independent variables, and whether they have opted in for a credit card as a dependent variable.

A model is then created to identify these clients (typically a logistic regression). Once the model is
trained it is then used to find other clients that are likely to take up a credit card – given you phone
them to market your product.

© 2019 - WorldQuant University – All rights reserved.

8
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

1.2 Regression

The second type of supervised learning relates to regression tasks. The difference between
classification and regression is that, for classification, the response variable – the potential output
of the algorithm – is discrete, whereas in regression tasks the response variable is continuous.

The most common example taught in machine learning classes for regression is a model for
predicting property prices. Almost every machine learning course will touch on it, and usually base
it on a multivariate linear regression. Hereby, the model is passed different features to treat as
independent variables – for example postal code, size in square meters, the number of floors, and
total rooms. Finally, the dependent variable is also passed to the model for training, which in this
case are the properties’ selling price.

We see a lot of examples of linear models in finance and econometrics. The Capital Asset Pricing
Model (CAPM) as well as Arbitrage Pricing Theory (APT) come to mind.

In fact, the Chartered Financial Analyst (CFA) Institute did a study that confirmed that linear
regression remains the primary workhorse for financial modeling (Zhu, Philpotts & Stevenson
(2012)). A benefit of this is that making use of non-linear models allows the user to harvest more
alpha and make more accurate forecasts than the majority of the competition – this is the skill set
we hope you pick up in this module.

Figure 1: Supervised learning pipeline (Adapted from HortonWorks)

© 2019 - WorldQuant University – All rights reserved.

9
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

As you can see from the chart above, we have a typical pipeline for a supervised learning
algorithm. Supervised learning is the machine learning task of learning a function that maps an
input, the feature vector, to an output based on an example of input-output pairs. So, we must
have labeled data.

As you can see in the pipeline, we have raw data; we then perform feature extraction/engineering
and combine it with the labeled data. Thereafter, we can use this data to train the model.

Feature engineering refers to the task of transforming your data, using domain knowledge, in
order to attempt to improve the performance of the machine learning algorithms. The model then
gets evaluated, based on its in-sample accuracy and loss, as well as its validation accuracy and loss,
then based on that we iterate by changing the model parameters and retrain the model.

When we want to score an out-of-sample set, we will give the model new data, perform the
feature engineering on it, provide it to the model as input features, and then the model will give us
the expected labels.

2 Unsupervised learning

The second class of machine learning we will look at is unsupervised learning. Unsupervised
learning refers to learning based only on input data and no corresponding output variables, or
labels. With unsupervised learning algorithms, the goal is for the system to generate its own model
of the underlying structure or distribution.

We use the name unsupervised learning, because there is neither a knowledgeable expert to label
data nor correct answers given to the system. In unsupervised learning, “algorithms are left to
their own devices to discover and present the interesting structure in the data” (Brownlee, 2016).
Unsupervised learning is also split into two subclasses – namely, clustering and representation
learning.

2.1 Clustering

Grouping similar datasets together helps the user to discover interesting structures. Our task is to
assign each observation to an appropriate class, more commonly referred to in this setting as
‘clusters’.

Clustering algorithms are commonly used for client segmentation to uncover hidden groupings for
anomaly detection and image compression.

© 2019 - WorldQuant University – All rights reserved.

10
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

Examples of clustering are found in the field of medicine. Breast cancer can be better detected
based on images in which certain parts of tissue will light up based on their features, aiding in the
identification of tumors and abnormalities. Another example, in genetics, is the hierarchical
clustering used in gene expression.

2.2 Representation Learning

“Representation learning is a set of techniques that allows a system to automatically discover the
representations needed for feature detection or classification from raw data. This replaces manual
feature engineering and allows a machine to both learn the features and use them to perform a
specific task.” – Bengio, Courville & Vincent (2014)

This is where algorithms, like Principal Component Analysis, Linear Discriminant Analysis, and T-
distributed Stochastic Neighbor Embedding algorithms, come into play. They generally act to help
extract features which form part of feature engineering.

Generally, they try to reduce large feature space (dimensionality reduction). This is popular within
factor models where analysts will pass through many features that they think are important for
predicting an asset, like economic data and zero investment portfolios (which act as a proxy for
financial accounting ratios). They may then use principal component analysis or some other
dimensionality reduction technique to reduce it to fewer features, but those features are very
information rich.

© 2019 - WorldQuant University – All rights reserved.

11
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

Figure 2: Principal component analysis is used for image compression

3 Reinforcement learning

Our third class of machine learning is reinforcement learning, where the algorithms are “learning
what to do — how to map situations to actions — so as to maximize a numerical reward signal.” –
Sutton & Barto (1998)

In reinforcement learning, the algorithm is not taught which actions to take when it encounters a
certain situation. Instead, the algorithm is given a situation and must apply different actions to it,
each of which yields a numerical reward. These different actions yield multiple rewards and the
algorithm’s implicit task is to find the action which maximizes its total reward. However, these are
not always linked to isolated situations as, in some cases, actions may affect consequent
situations, and therefore, the potential rewards.

According to Sutton and Barto (1998) these are the two distinguishing characteristics of
reinforcement learning: trial-and-error search and delayed reward. Reinforcement learning is
especially useful for the optimization of strategies for tasks and learning objectives from behavior.

A very common example of the optimization of a strategy for a task is AI applied to video games
(for instance, with Atari Games, video games in which you traditionally play as an agent – say,

© 2019 - WorldQuant University – All rights reserved.

12
MScFE 650 Machine Learning in Finance - Module 1: Unit 1

Mario from the Mario Brothers). Here you need to take certain actions like jumping to acquire
coins or to avoid in-game monsters. Atari Games can be trained using the reinforcement learning
paradigm. Another game-based example is from OpenAI, which released a bot for Dota 2 that is
capable of beating world-class players (follow this link for a video example).

In the context of finance, reinforcement learning can be used for portfolio management, although
this is a complex problem outside of our current scope. An interesting paper you can read is “A
Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem”.

© 2019 - WorldQuant University – All rights reserved.

13
MScFE 650 Machine Learning in Finance - Module 1: Unit 2

Unit 2: The Community, Projects, Events and Resources

Assorted Links and Projects

In this set of notes, we are going to take a look at some of the projects, events, and resources that
are part of the machine learning community. These are especially useful to anyone interested in
getting started with machine learning.

Open-source projects

The open-source machine learning projects we’ll be looking at are all hosted on GitHub, and
include OpenPose, Person Blocker, and an unnamed globally and locally consistent image
completion algorithm. For the finance projects, we add Zipline and PGPortfolio. To begin, let’s look
at OpenPose.

OpenPose is an open-source algorithm that, in real time, can detect a human’s body, hand, and
facial key points in single images. It allows users to input an image, a video, or a camera feed, and it
not only detects whether people are present but maps the key points of their body to create a 3D
model. Although it may at first seem purely for fun, this type of algorithm can be used in practical
and commercial applications such as body-detection for security systems, and body-capturing for
gaming such as the Wii.

Figure 3: OpenPose body and facial recognition

Person Blocker, on the other hand, is another machine learning algorithm which uses a region-
based convolutional neural network to analyze images. However, instead of mapping a person’s

© 2019 - WorldQuant University – All rights reserved.

14
MScFE 650 Machine Learning in Finance - Module 1: Unit 2

key points like OpenPose, Person Blocker detects certain objects in the image and replaces it with
a region of ‘video noise’ (or static). This can be used to block a person out of the picture, and a
model could even be trained to learn what ‘Kenny’ looks like and in every photo crop out ‘Kenny’.

Figure 4: Image analysis and replacement by Person Blocker

Our third example is of a globally and locally consistent image completion algorithm, which allows
users to erase parts of an image which are then replaced. The algorithm analyzes the partial image
to figure out what it should look like and replaces the empty space with another consistent image.

Figure 5: Examples of image replacement by machine learning algorithm

© 2019 - WorldQuant University – All rights reserved.

15
MScFE 650 Machine Learning in Finance - Module 1: Unit 2

Despite the seemingly limited applicability of these three algorithms in finance, computer vision
nonetheless has been put to use in financial contexts. For instance, optical character recognition
(OCR) is used on bank statements, receipts, and invoices to digitally record any printed text.
Another example is of photos of oil tankers. The photos are used to determine how much supply is
available to match the underlying demand for oil.

Moving on to the finance examples, we chose two open-source projects: Zipline and PGPortfolio.
While Zipline is not an example of machine learning, it is a very powerful tool for setting up a
research environment for quantitative finance projects.

Zipline is a Pythonic algorithmic trading library. It is an event-driven system for backtesting,

currently used in production as the backtesting and live-trading engine powering Quantopian – a
free, community-centered, hosted platform for building and executing trading strategies.

Figure 6: Zipline

In relation to PGPortfolio, the paper “A Deep Reinforcement Learning Framework for the
Financial Portfolio Management Problem” explains clearly how reinforcement learning can be
used in managing a portfolio of stocks. The paper goes on to show how the suggested framework
outperforms other online learning techniques for portfolio management. It is also worth noting
that the author implemented it on a portfolio of cryptocurrencies.

© 2019 - WorldQuant University – All rights reserved.

16
MScFE 650 Machine Learning in Finance - Module 1: Unit 2

Figure 7: Back Test 3 – 2017-03-07-4:00 to 2017-04-27-8:00 (UCT), log-scale accumulated wealth.

All algorithms struggle and consolidate at the beginning of this experiment, and both of the EIIEs
experience two major dips on March 15 and April 9. This diving contributes to their high Maximum
Drawdown in the text. Nevertheless, this is the best month for both EIIEs in terms of final wealth.
(Adapted from Jiang, Xu and Lang, 2017)

If you are interested in finding similar projects, GitHub is an ideal place for other open-source
machine learning projects. The computer vision examples were sourced from the following link.

Open-source frameworks

A framework in the context of machine learning refers to any set of tools, which may include an
interface or library, that supports developers and allows us to quickly build machine learning
models. Frameworks make this possible as they can allow us to circumvent difficult programming
challenges and simplify the process significantly.

There are many popular open-source frameworks in Python such as scikit-learn, Keras,
TensorFlow, and Spark ML, to mention a few.

At this stage, exploring scikit-learn, Keras, and TensorFlow in depth is beyond the scope of
introducing you to machine learning. We’ll examine scikit-learn to showcase what these open-

© 2019 - WorldQuant University – All rights reserved.

17
MScFE 650 Machine Learning in Finance - Module 1: Unit 2

source packages offer us as developers. Later in the course, you will be introduced to Keras and
TensorFlow.1

Scikit-learn

Scikit-learn is a free machine learning library, and is essentially a simple and efficient tool used for
data mining and data analysis. The library includes a variety of algorithms used for supervised
learning, unsupervised learning, model selection and evaluation, and dataset transformations,
among others (scikit-learn user-guide, 2017).

The scikit-learn library is accessible to everybody and can be used in various contexts. On the
technical side, the library is predominantly written in Python, and built on NumPy, SciPy, and
matplotlib. It is an open-source, commercially usable library and, since it has a BSD license, scikit-
learn could even be incorporated in proprietary products.

Scikit-learn not only includes a library of tools, but also some datasets for developers to play
around with and test its built-in algorithms. For an example, let’s look at the image below, where
we can see the results of clustering algorithms applied to different datasets. The image shows us
how well different algorithms perform with respect to clustering. In the first row, you’ll notice a
smaller circle within a larger circle, which depicts how well the different algorithms do at
clustering the data. In the fourth column (under SpectralClustering), where the inner circle is
orange and the outer circle is blue, we can see what the output from an effective algorithm should
look like.

For the other types of patterns, it is clear that spectral clustering performs well. Our brains
achieve the same results automatically.

1 We will be focused on Python-based frameworks in this course. It is strongly recommended to use Python
for machine learning, as it has become the de facto language for machine learning and offers many
advantages over alternatives.

© 2019 - WorldQuant University – All rights reserved.

18
MScFE 650 Machine Learning in Finance - Module 1: Unit 2

Figure 8: Results of clustering algorithms applied to different datasets

Now, let’s look at some of the modeling tools included in scikit-learn. In this example we’ve
imported a linear regression model and a function that allows us to split the data between training
and testing sets.

from sklearn.linear_model import LinearRegression

from sklearn.linear_selection import train_test_split
import mglearn

Once we’ve imported the model and functions, we can generate our own data, and use them in
combination. After the data has been generated in the first line, random sampling is performed to
sample 60 observations, which are then split into training and test data.

X, y = mglearn.datasets.make_wave(n_samples=60)

X_train, X_test, y_train, y_test = train_test_split (X, y,

random_state+42)

Thereafter, a linear regression model is fitted.

lr = LinearRegression().fit(X_train, y_tran)

© 2019 - WorldQuant University – All rights reserved.

19
MScFE 650 Machine Learning in Finance - Module 1: Unit 2

Figure 9: Linear regression

Finally, the coefficient, the intercept, and how well the model scored on in-sample (training set
score) and out-of-sample (test set score) data can be printed out.

print(“lr.coef_: {}”.format(lr.coef_))
print(“lr.intercept_: {}”.format(lr.intercept_))
print(“Training set score: {:.2f}”.format(lr.score(X_train,
y_train)))
print(“Test set score: {:.2f}”.format(lr.score(X_test, y_test)))

This returns:
lr.coef_: [ 0.394]
lr.intercept_: -0.031804343026759746
Training set score: 0.67
Test set score: 0.66

The take-away from these examples should be that scikit-learn allows us to fit a model to the data
in six lines of code. Rather than having to create everything from scratch, open-source libraries
like scikit-learn, Keras, and TensorFlow are immensely helpful, and help speed up the process by
circumventing the need to code similar models and functions from first principles.

© 2019 - WorldQuant University – All rights reserved.

20
MScFE 650 Machine Learning in Finance - Module 1: Unit 2

Research groups

Once you have completed your MSc in Financial Engineering, many of you will perhaps want to
consider doing a PhD program or joining a research group. Here are a few of the most popular
options:

o Machine Learning Group, University of Toronto

o MILA Lab

o CILVR at NYU

o Stanford Machine Learning Group

o Oxford Deep Learning Group

o Google Research

o Google DeepMind

o OpenAI

o Facebook AI Research (FAIR)

o Microsoft Research

Other online courses

One of the major objectives of this course is to help you develop skills to help yourself. This means
that you will frequently be prompted to explore the internet and other resources. For those who
wish to go beyond this seven-week course and explore machine learning in more depth, some
highly recommended courses include:

Machine Learning: Introduction (Stanford University), taught by Andrew Ng, which is where
enthusiasts usually start. From that point you can decide on two things either to learn more about
machine learning in general – in which case there is the Machine Learning Specialization
(University of Washington) course on Coursera – or for those who are interested in deep learning,
there is a specialization dedicated to the topic also taught by Andrew Ng via an organization called
Deep Learning Specialization (Deeplearning.ai).

And finally, to expand on many of the topics covered in this course, there is also Machine Learning
for Finance (New York University). This program has several examples of reinforcement learning.

© 2019 - WorldQuant University – All rights reserved.

21
MScFE 650 Machine Learning in Finance - Module 1: Unit 3

Unit 3: Machine Learning in Financial Contexts

To continue developing our understanding of how machine learning applies to finance, we will look
at some of the ways different machine learning algorithms are used.

An important point to note is that the majority of academic papers on the topic of alpha design
with machine learning focus on using classification techniques rather than regression. Regression
techniques are commonly used in econometric models such as ARIMA (discussed in Module 3 of
Econometrics) and GARCH (discussed in Module 4 of Econometrics). In the classification setting, it
is typical to see the author predict the direction of a next move above some threshold, such as the
daily volatility. Accordingly, if the stock goes above threshold it is marked as ‘1’ and if it falls below
the threshold it is marked as ‘–1’ or a ‘0’ (depending on how your model is structured).

Regression
The following are some examples of regression, used in finance:

1 Forecasting ZIPs: These are zero investment portfolios. One of the advantages about using a
zero-investment portfolio is that you can represent financial accounting ratios as an
engineered feature. You can read more about this in the book: Quantitative Equity Portfolio
Management.
Fama and French also use the idea in their famous paper “Common Risk Factors in the Returns
on Stocks and Bonds” (1993). What they found is that expected returns could be well
explained by three factors – namely, market, size, and value.

2 The Capital Asset Pricing Model and Arbitrage Pricing Theory are based on linear regression.

3 Supporting decisions: Where analysts would typically fit a univariate or multivariate linear
regression to a dataset, they would then, on seeing how well the factor loading explains it,
decide to use it in their decision-making process. A good example of this can be found in the
book: An Introduction to Statistical Learning. On page 18, there is an example of regression in
which income as a function of education and seniority is analyzed.

© 2019 - WorldQuant University – All rights reserved.

22
MScFE 650 Machine Learning in Finance - Module 1: Unit 3

Figure 12: The plot displays income as a function of years of education and seniority in the Income dataset.
The purple surface represents the true underlying relationship between income and years of education and
seniority, which is known since the data are simulated. The orange dots indicate the observed values of
these quantities of 30 individuals (Adapted from Hastie, Tibshirani and Friedman, 2009).

4 Pricing interest rates.

5 Forecasting accounting ratios like Earning per Share (EPS):

Neural network earnings per share forecasting models: A comparative analysis of alternative
methods.

Classification

On the classification side, we see the following examples:

1 Predicting the direction of the next move for a portfolio or individual stocks
One such paper, by Matthew Dixon, Diego Klabjan, and Jin Hoon Bang, titled “ Classification-
based Financial Markets Prediction using Deep Neural Networks”, provides a good basis.

2 Propensity modeling for targeted marketing

This technique commonly uses logistic regression to identify which clients have a high
probability of taking up a specific product, should the firm market to them. An example is
scoring clients in your database to see who will likely take up a credit card.

© 2019 - WorldQuant University – All rights reserved.

23
MScFE 650 Machine Learning in Finance - Module 1: Unit 3

3 Anti-money laundering
This is used in retail banking to identify clients who are taking part in illegal activities.

4 Credit scoring
In countries like South Africa, retail banks have their own credit-scoring models that take in a
range of features about a client and then predict the likelihood of that client repaying the loan.
These models are also used to set interest rates for clients.

5 Regime detection
A market can be in various regimes throughout a given period. Classification models can be
used to identify which market regime we are in – for example, a trending or mean reverting
regime, and then analysts can apply alpha models which will best exploit it.

Clustering

Within the class of unsupervised learning, clustering models are used for:

1 Client segmentation
It is typical to apply client segmentation when trying to model clients using a linear model like
logistic regression. This allows the analyst to fit a different model to each of the different
clusters.

2 Regime clustering
Unsupervised learning can also be used to identify which regime is currently active.

3 Document retrieval
A common example of this is when one has a document or a piece of information, such as a
Stock Exchange News Service announcement, but seeks other resources like it.

Representation learning

For representation learning, we see:

© 2019 - WorldQuant University – All rights reserved.

24
MScFE 650 Machine Learning in Finance - Module 1: Unit 3

1 Factor models using PCA

A good example is found on Quantopian: Using PCA to trade a Long/Short Basket.

2 Identifying pairs for statistical arbitrage

For more on this, see Pairs Trading with Machine Learning.

Reinforcement learning

3 Portfolio management
For more on this, see Policy Gradient Portfolio: A Deep Reinforcement Learning Framework
for the Financial Portfolio Management Problem.

4 Automated trading
This is specifically on the sell-side, or you could market it as generating execution alpha. So,
how can you spend less on transaction fees or how could you get a better price, closer to the
volume-weighted average price?

5 Options Pricing
For more on this, see A Reinforcement Learning Approach for Pricing Derivatives.

Funds that use machine learning

The above examples not only reflect how machine learning can be used; they are also widely used
in industry. Various entities in the financial sphere have incorporated machine learning into their
practices. Let’s take a look at six different funds that are known to use machine learning.

Bridgewater Associates

Bridgewater is famous for being the largest hedge fund in the world – using a systematic global
macro approach, Ray Dalio, the founder of Bridgewater, mentions in his book, Principals: Life and
Work, how they use AI to aid in decision support.

Citadel

Citadel is an American global financial institution, founded in 1990 by Kenneth Griffin in his
dormitory at Harvard. It is one of the flagship success stories for quantitative funds. By 2015 it

© 2019 - WorldQuant University – All rights reserved.

25
MScFE 650 Machine Learning in Finance - Module 1: Unit 3

had claimed its place as one of the world’s largest alternative investment firms (follow this link for
more information).

Two Sigma

Two Sigma, a hedge fund based in New York city is famous for implementing the scientific method
to investing. The “About Us” page on its website reads: “We apply the scientific method to
investment management. Some people call us systematic. Here’s what that means to us: We
combine massive amounts of data, world-class computing power and financial expertise to
develop sophisticated trading models. And we use technology in an effort to constantly refine our
investment process and our insights into how the markets will behave. Risk is a living thing in our
business, and we watch over it carefully” (follow this link for more information).

Man AHL

Man AHL is a diversified quantitative investment manager that has been a pioneer in the
application of systematic trading since 1987. The following link is from their website, explaining
how they have made use of machine learning algorithms for years.

A section on their website reads: “Man AHL has been conducting research in this area for decades.
However, until quite recently the results of the research, which featured buzzwords in the 1990s
such as ‘genetic algorithms’ and ‘neural networks’, floundered against the requirements for raw
computational power and data density. With significant advances in these areas, as well as
improvements in underlying theory, we have now been actively applying machine learning, in
various forms, within some of our less trend-focused client programs since 2014. This year, a suite
of machine learning algorithms, constrained to behave in a predominantly momentum-like
manner, was incorporated into Man AHL’s flagship trend-following strategies.”

D.E. Shaw & Co.

Computer scientist and investor David Elliot Shaw founded D.E. Shaw & Co. He is most famous for
being part of the team at Morgan Stanley, 1985, that discovered pairs trading (a form of statistical
arbitrage).

As of January 2018, the firm manages more than 47 billion dollars and has offices in North
America, Europe and Asia.

© 2019 - WorldQuant University – All rights reserved.

26
MScFE 650 Machine Learning in Finance - Module 1: Unit 3

Point72

Point72 is a global asset management firm led by Steven Cohen. It uses Discretionary Long/Short,
Macro and Systematic strategies to invest in eight offices across the globe (follow this link for
more information).

© 2019 - WorldQuant University – All rights reserved.

27
MScFE 650 Machine Learning in Finance - Module 1: Unit 4

Unit 4: Academic Papers in Review

This paper is publicly available for students, via ResearchGate.

Notes

Every year there is a competition hosted in which models are compared in the context of business
time series problems. Different algorithms are used to establish which yields the best
performance. Different business-type datasets are also used, each exhibiting certain features. In
the M3 Competition there were 3003 time series. Here is a link to the M4 Competition, which
includes 100,000 time series.

There are very few large-scale comparison studies for machine learning regression techniques
applied to time series. You will recall, from earlier in the module, that the majority of papers
related to financial applications of machine learning are fixated on using classification techniques,
predicting if stock is going to go up or down.

The two best algorithms reported in “An Empirical Comparison of Machine Learning Models for
Time Series Forecasting” are neural networks – the authors refer to these as a multilayer
perceptron (MLP) – and Gaussian Processes. They also test different preprocessing methods and
how they affect performance. There have been many studies comparing the results of neural
networks against traditional econometric models for forecasting and other econometric problems.

Importantly, the results of these studies are somewhat mixed, but the overall consensus is that
neural networks tend to outperform classical linear techniques. In this particular case, neural
networks perform best, followed by Gaussian processes. Note, however, that Gaussian processes
run much faster than neural networks. The table below presents the computation time for every
method per time series. What we can see here is that Gaussian processes are relatively fast.

MLP 1.6076
BNN 2.8251
RBF 0.6917
GRNN 0.0885

© 2019 - WorldQuant University – All rights reserved.

28
MScFE 650 Machine Learning in Finance - Module 1: Unit 4

KNN 0.0068
CART 0.0803
SVR 0.3099
GP 0.6268

Table 1: The computation time for every method per time series (From Ahmed et al., 2010)

Another important aspect to note is that although SVMs are one of the top models in
classification, in regression they do not seem to maintain the top spot. In the second paper
reviewed in these notes, we will see that SVMs are coming up again as a top performer.

In this course, we are going to focus on deep learning models (neural networks) rather than SVMs,
because deep learning is setting the benchmarks in almost all applications. In finance SVMs,
however, Random Forests are popular, and many papers also include neural networks. A neural
network is a useful algorithm to learn. Later, we will give this special attention.

Trends and seasonality affect forecasts negatively. It is easier to forecast a stationary time series.
The table below reveals the different error rates for trends, no trends, seasonal, and non-seasonal.

Model Trend No-Trend Seasonal Non-Seasonal Zero-Hid Non-Zero-Hid

MLP 0.1366 0.0597 0.1081 0.0811 0.0848 0.0869
BNN 0.1634 0.0716 0.1978 0.0831 0.1078 0.0937
RBF 0.2109 0.0803 0.1511 0.1191 0.1304 0.114
GRNN 0.1729 0.0689 0.1169 0.1015 0.1087 0.0958
KNN 0.1724 0.0683 0.1154 0.1011 0.1084 0.0944
CART 0.2043 0.0776 0.1354 0.1175 0.126 0.1104
SVR 0.1655 0.0658 0.1113 0.0971 0.1029 0.0935
GP 0.1565 0.0631 0.1126 0.0911 0.0981 0.0888

Table 2: The overall summetric MAPE (SMAPE-TOT) by other categories on all the 1045 time
series for the LAGGED-VAL preprocessing method (From Ahmed et al., 2010)
Finally, “An Empirical Comparison of Machine Learning Models for of Time Series Forecasting”
looks at the typical transformations performed on time series data, as well as three new
transformations. The typical transformations performed in the data for these types of problems,
performed in this order, are:

1 Log transformation: We have to remember that prices are often assumed to be log-normally
distributed and thus returns are normally distributed.
2 Deseasonalization
3 Scaling

© 2019 - WorldQuant University – All rights reserved.

29
MScFE 650 Machine Learning in Finance - Module 1: Unit 4

The additional methods considered in the paper are as follows:

1 Lagged-values, which is no special preprocessing. This is where the input feature vector to
your model are the features at 𝑡 – 1, where your target is at 𝑡.
2 Time series differencing, which is what you are used to in the ARIMA style models – it is the
first backward difference and then you apply the model on the differences.
3 Taking a moving average, mainly to eradicate noise. Financial time series are infamous for
having a very low signal to noise ratio. When using a moving the average you will want to have
several moving averages with different window periods, to preserve different levels of detail.

Results

The result of the authors’ study is that the MLP / neural net outperforms the other models;
however, the “superiority of MLP is genuine and not due to ability to switch to a linear model”.

Differencing the time series resulted in a much worse performance. This is very important. Dr
Lopéz de Prado brings this up in his book regarding integer differentiation. Here it comes up again
and we see that differencing time series resulted in a much worse performance. A possible
explanation as to why it did worse is that the absolute level of the time series contributes
important information. For example, there may be a mean-reverting behavior in the series that
indicates when a series is more likely to change direction or reverse when a high or low level is
reached.

While differencing for linear models is a well-understood operation, for which tests exist to notify
practitioners of its need, the above is not true for non-linear models like neural networks. Dr
Lopéz de Prado suggests an elegant solution that is termed ‘fractional differentiated features’. He
goes into detail about the problem of memory versus stationarity, which you can read more about
in Chapter 5 of his textbook.

Next, we will go over some graphs from the article.

© 2019 - WorldQuant University – All rights reserved.

30
MScFE 650 Machine Learning in Finance - Module 1: Unit 4

Figure 13: The average ranks with 95% confidence limits for the multiple comparison with the best
test for the LAGGED-VAL preprocessing method. The dashed line indicates that any method with
confidence interval above this line is significantly worse than best.
(Adapted from Ahmed et al., 2010)

The table below presents their error measure. MLP is the lowest, followed by Gaussian processes,
but when one turns to Table 9 for computation time, it is clear to see that MLP is far more
computationally expensive than the Gaussian process.

Model SMAPE
MLP 0.0384
BNN 0.0411
RBF 0.0442
GRNN 0.0513
KNN 0.0511
CART 0.0566
SVR 0.0468
GP 0.0397

Table 3: SMAPE for the eight models for the time series example
(From Ahmed et al., 2010)

© 2019 - WorldQuant University – All rights reserved.

31
MScFE 650 Machine Learning in Finance - Module 1: Unit 4

Note that the error measure that they are using, SMAPE, refers to the symmetric mean absolute
percentage error, and because it is a relative measure the values across datasets can be combined
into one overall figure.

This paper evaluates multiple classifiers for stock price direction prediction, but in particular
ensemble techniques.

Notes

This paper looks ahead, one year into the future. Naturally, a lot can change over this of time. The
authors found that, for the purposes of their prediction, the top algorithms were Random Forest
and SVMs.

It is worthwhile noting that the authors make a distinction between the linear methods we saw in
Econometrics and the nonlinear methods found in machine learning. They particularly note that:
“Although EMH is generally accepted, it was initially based on traditional linear statistical
algorithms. Many researchers have already rejected the hypothesis by using algorithms that can
model more complex dynamics of the financial system” (Ballings et al., 2015).

In the literature review of the papers it is clear that the majority of stock price direction prediction
literature has focused on Logistic Regression, Neural Networks, K-Nearest Neighbors and SVMs.
Neural Networks are by far the most common followed by SVMs and logistic regression.

Prediction method
LR NN KN SVM AB RF KF
Schönenburg (1990) X
Bessembinder and Chan (1995) X
Brownstone (1996) X X
Saad, Prokhorov, and Wunsch (1996) X
Kim and Chun (1998) X
Saad, Prokhorov, and Wunsch (1998) X
Kim and Han (2000) X
Kuo et al. (2001) X
Kim (2003) X X
Kim and Lee (2004) X
Rodriguez and Rodriguez (2004) X X X X
Huang et al. (2005) X X
Kumar and Thenmozhi (2005) X X X X
Lunga and Marwala (2006) X
Wu et al. (2006)

© 2019 - WorldQuant University – All rights reserved.

32
MScFE 650 Machine Learning in Finance - Module 1: Unit 4

Wang and Chan (2007) X

Huang et al. (2008) X X X X
Senol and Ozturan (2008) X X
Lai, Fan, and Chang (2009)
Lee (2009) X X
Ou and Wang (2009) X X X X
Kara, Boyaciogly and Baykan (2011) X X
Wei and Cheng (2011) X
Subha and Nambi (2012) X X
Lin, Guo, and Hu (2013) X
De Oliveira, Nobre, and Zárate (2013) X
Chen, Chen, Fan, and Huang (2013) X
Rechenthin et al. (2013) X X X
Ji, Che, and Zong (2014) X
Bisoi and Dash (2014) X
Zikowski (2015) X
Hafexi, Shahrabi, and Hadavandi
(2015) X
Patel et al. (2015) X X X
This study X X X X X X X

Table 4: Algorithms for stock price direction prediction used in literature

(Adapted from Ballings et al., 2015)

Interestingly, SVMs will continually come up as a best-performing algorithm for financial time
series. In the figure below we can see the median area under the curve (AUC). What is interesting
is that neural nets did not perform as well as the other classifiers, such as Random Forest and
SVMs.

Figure 14: 5x2fc median AUC per algorithm

(Adapted from Ballings et al., 2015)

© 2019 - WorldQuant University – All rights reserved.

33
MScFE 650 Machine Learning in Finance - Module 1: Bibliography

Bibliography

Recommended Reading

Ahmed, N. K., Atiya, A., El Gayar, N., and El-Shishiny, N. (2010). ‘An Empirical Comparison of
Machine Learning Models for Time Series Forecasting’. Econometrics Reviews, 29(5-6), p. 594.
DOI: 10.1080/07474938.2010.481556

Chincarini, L. B., and Kim, D. (2006). Quantitative Equity Portfolio Management: An Active
Approach to Portfolio Construction and Management. McGraw-Hill.

Ballings, M., Van den Poel, D., Hespeels, N., and Gryp, R. (2015). ‘Evaluating Multiple Classifiers for
Stock Price Direction Prediction’. Expert Systems with Applications, 42(20), p. 7046. DOI:
10.1016/j.eswa.2015.05.013

Zhang, W., Cao, Q., and Schniederjans, M. J. (2004). ‘Neural Network Earnings per Share
Forecasting Models: A Comparative Analysis of Alternative Methods’. Decision Sciences, 35(2), p.
205. DOI: 10.1111/j.00117315.2004.02674.

References

Bengio, Y., Courville, A. and Vincent, P., 2013. ‘Representation learning: A review and new
perspectives’. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8), p. 1798.

Brownlee, J. (2016). Supervised and Unsupervised Machine Learning Algorithms. [online]

Available at: https://machinelearningmastery.com/supervised-and-unsupervised-machine-
learning-algorithms/

Lopéz de Prado, M. (2018). Advances in Financial Machine Learning. Wiley Publishers.

McCarthy, J., Minsky, M.L., Rochester, N. and Shannon, C.E. (1955). A Proposal for the Dartmouth
Summer Research Project on Artificial Intelligence, Dartmouth proposal.

Moravec, H. P. (1988). Mind Children: The Future of Robot and Human Intelligence. Harvard
University Press.

© 2019 - WorldQuant University – All rights reserved.

34
MScFE 650 Machine Learning in Finance - Module 1: Bibliography

Scikit-learn. (2017). Scikit-learn User Guide. [online], Available at: http://scikit-

learn.org/stable/user_guide.html

Searle, J.R. (1980). ‘Minds, Brains, and Programs’. Behavioral and Brain Sciences, 3(3), p. 417.

Sutton, R.S. and Barto, A.G. (1998). Reinforcement Learning: An Introduction. MIT press.

Trefethen, L.N. and Bau, D. (1997). Numerical Linear Algebra, SIAM.

Turing A.M. (1950). ‘Computing Machinery and Intelligence’. Mind, 49, p. 433.

Zhu, M., Philpotts, D. and Stevenson, M. J. (2012). ‘The Benefits of Tree-Based Models for Stock
Selection’. Journal of Asset Management, 13(6), p. 437. DOI: 10.1057/jam.2012.17

© 2019 - WorldQuant University – All rights reserved.

35
MScFE 650 Machine Learning in Finance - Module 1: Collaborative Review Task

Collaborative Review Task

The collaborative review task is designed to test your ability to apply and analyze the knowledge
you have learned in the module.

Instructions
o Work through the notebook, answer all questions/problems.

o You are allowed to consult the internet, as well as your peers on the forum.

o Your answers and solutions to the problems should be added to this notebook.

o Submit your final work as an html file.

o Note that Python (version 3.6.4) has been used to calculate the solutions.

o Hint: In this assignment, you work with a single linear regression problem where you
generate data yourself. This means that you have an idea of what the computed solution
should look like.

1 Regression

1.1 We will start with the simplest possible regression problem. Solve the following system of
linear equations:

𝒙 = 1.0
𝒙 = 0.5
𝒙 = 0.7

1.1.1 Write the system above as a linear system of equation of the form 𝐴𝑥 = 𝑏.
(3 marks)

1.1.2 Write down the normal equations, 𝐴𝑇𝐴𝑥 = 𝐴𝑇𝑏 for this system

(3 marks)

1.1.3 Solve the system of normal equations for the system above. Is there an intuitive
interpretation of the solution?

(3 marks)

© 2019 - WorldQuant University – All rights reserved.

36
MScFE 650 Machine Learning in Finance - Module 1: Collaborative Review Task

1.2 We will now revisit the well-known linear regression problem – first formulating it as the
solution of an over-determined system, then as an optimization problem. For solving this
example, we will use the computer.

We are given data in the form of (𝑥, 𝑦) pairs and we know that, given 𝑥, 𝑦 = 𝑚𝑥 + 𝑐. Since one
can fit a straight line through two points: we only need two pairs to fit a line. But there is a rub:
the 𝑦 values are contaminated by noise (at this point you may want to look at the example
below), and it is not possible to find a single line that will pass through all points. We therefore
need to find the line that best fits the given values. We will lead you through the solution.

Given (𝒙𝒏 , 𝒚𝒏 ), 𝒏 = 1, . . . , 𝑵, we need to solve for 𝒎 and 𝒃 from 𝒎𝒙𝒏 + 𝒃 = 𝒚𝒏 , 𝒏 = 1, . . . , 𝑵.

1.2.1 Write this as a linear system where the unknowns are 𝒎 and 𝒃.

(3 marks)

1.2.2 Write down and solve the normal equations for the unknowns 𝒎 and 𝒃.

(3 marks)

1.3 Next, we treat this as an optimization problem. Instead of using the normal equations, we now
look for 𝑚 and 𝑏 that minimizes the mean squared error,
𝑁
1
𝐸 = ∑(𝑚𝑥𝑛 + 𝑏 − 𝑦𝑛 )2
𝑁
𝑛=1

∂𝐸 ∂𝐸
1.3.1 Write down the 𝑒𝑞𝑢𝑎𝑡𝑖𝑜𝑛𝑠 = 0 and = 0.
∂𝑚 ∂𝑏

(3 marks)

1.3.2 Write the equations as a linear system and solve for 𝑚 and 𝑏. How does your answer
compare with the answer obtained above?

(3 marks)

It should be clear that these are the same as the normal equations.

2 Solutions: using numpy

© 2019 - WorldQuant University – All rights reserved.

37
MScFE 650 Machine Learning in Finance - Module 1: Collaborative Review Task

2.1
In [1]:# This ensures that any graphs we generate will be part of
# the notebook, instead of opening in a separate window
%matplotlib inline

In [2]:import numpy as np # import the numerical linear algebra model

from matplotlib import pyplot as plt

In [3]:x_train = np.linspace(0,1,100)
y_train = 0.2*x_train+1+0.01*np.random.randn(x_train.shape[0])

plt.plot(x_train, y_train, ‘r.’)

plt.show()

2.1.1 Using numpy, create 𝐴and 𝑏 for the system 𝐴𝑥 = 𝑏 (here 𝑥 is the vector containing the
unknown weight 𝑤 and the offset 𝑏).
(5 marks)
2.1.2 Solve the normal equations for 𝒎 and 𝒃.
(5 marks)

2.1.3 Why do the numerical values differ from the values used to generate the data?
(3 marks)

2.2 Although it does not matter much for the simple example above, in general it is bad practice to
use the normal equations to solve the least-squares problem. Surprisingly enough, the culprit
is the covariance matrix 𝐴𝑇𝐴 . We cannot compute it in a numerically stable fashion.
Technically, its condition number is roughly the square of the condition number of 𝐴 and, in
practice, this means that we lose roughly half the precision if we calculate the covariance
matrix.

© 2019 - WorldQuant University – All rights reserved.

38
MScFE 650 Machine Learning in Finance - Module 1: Collaborative Review Task

A full discussion is beyond the scope of this course, but the interested reader is encouraged to
consult Trefethen & Bau (1997).

A numerically stable way to solve the linear least-squares problem is to first do a 𝑸𝑹

factorization of 𝑨. If 𝑨 is an 𝒎 × 𝒏 dimensional array with 𝒎 > 𝒏 (more equations than
unknowns), then the 𝑸𝑹 factorization is given by

𝑨 = 𝑸𝑹

where 𝑸 is an 𝒎 × 𝒏 matrix with orthogonal columns, i.e. 𝑸𝑻 𝑸 = 𝑰, and 𝑹 is an 𝒏 × 𝒏 upper

triangular matrix. Assuming that 𝑨 has full rank, i.e. its columns are linearly independent, then
it can be shown that 𝑹 is non-singular. One has to be careful, but there are a number of
numerically stable algorithms to compute the 𝑸𝑹 factorization.

Let's now substitute the 𝑨 = 𝑸𝑹 in the normal equations:

(𝑸𝑹)𝑻 𝑸𝑹𝒙 = (𝑸𝑹)𝑻 𝒃.

Taking the orthogonality of 𝑸 and the non-singularity of 𝑹 into account, this equation boils
down to

𝑹𝒙 = 𝑸𝑻 𝒃.

Since 𝑹 is non-singular and upper triangular, this system is easily and efficiently solved.

What has happened here? After we calculated the 𝑸𝑹 algorithm in a stable manner, we did
many of the calculations analytically (using 𝑸𝑻 𝑸 = 𝑰 and cancelling 𝑹𝑻 from both sides), and
that is done with infinite precision.

2.2.1 Using numpy, create 𝑨and 𝒃 for the system 𝑨𝒙 = 𝒃 (here 𝒙 is the vector containing the
unknown weight 𝒘 and the offset 𝒃).
(3 marks)

In [4]: # Your code here

# End of your code

print(Q.shape, ‘\n’)
print(‘R = ‘, R, ‘\n’)
print(‘Q^TQ = ‘, np.dot(Q.T,Q))
print(m, b)

© 2019 - WorldQuant University – All rights reserved.

39
MScFE 650 Machine Learning in Finance - Module 1: Collaborative Review Task

3 Solutions: using scikit-learn and Keras

3.1 First, we need to import the necessary module from scikit-learn:

In [7]: # Import modules from sklearn
from sklearn.linear_model import LinearRegression

3.1.1 Solve the linear least-squares problem using scikit-learn:

(5 marks)

In [6]: # Fit the data to the model

# Print the weight and offset. The values are the same as
from # the normal equations

3.2 Next, import the necessary modules from Keras:

In [8]: from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import Adam

3.2.1 Solve the linear least-squares problem using Keras.

(5 marks)
3.2.2 Use the Adam optimizer and experiment with different learning rates (Note: You should
find everything you need on the internet).
(5 marks)

In [9]: # Construct and fit your model here

# Use a batch_size=20, epoch=300

In [10]: # Print model weights, might not be exactly the same as

above.

© 2019 - WorldQuant University – All rights reserved.

40