Machine learning for the

prediction of stock returns

– For professional investor use only –

05/20
At a glance
· Once associated with science fiction, machine learning has come a long way over the
past decade and is now integrated in everyday life. At LOIM, we jumped on the machine
learning train early on. It was not an easy journey and we share our experience and
views in this note.
May 2020
· Implementing quantitative models in asset portfolios is the bread and butter
of systematic investments, and machine learning can help with detecting more
complex predictive relationships beyond traditional factors.
· However, the application of machine learning is far from straightforward, and especially
in finance. There are several challenges to overcome, from the forward-looking bias to
discovering spurious relationships. We believe that only a combination of strong
financial expertise with a good knowledge of the data science will produce a strategy
Machine learning is able that performs not only on the paper.
to discover more complex · Our research efforts did bear fruit. We are pleased to announce a new investible
predictive relationships beyond long-short equity strategy as a building block of ARP offering. The strategy seeks
those used in simplistic to achieve the dual objective of generating performance and promoting responsible
factor-based models. companies through application of ML techniques to traditional fundamental and
technical stock characteristics, as well as ESG data.

p.05
For professional investor use only. Please see
important information at the end of this document. · Introduction  p.2
This document has been prepared by:
· Systematic investment without
Alexey Medvedev
Portfolio Manager, Systematic Equities
human biases  p.2
Laurent Joué · Machines will not work without human  p.2
Senior Portfolio Manager, Commodities and
Alternative Risk Premia Strategies · How do they learn?  p.4
Combining strong financial
Marc Pellaud expertise with a good knowledge
Senior Portfolio Manager, Commodities and · LOIM-ML model: new alpha or
Alternative Risk Premia Strategies
old factors?  p.5
of data science.
Cesar Moura
Senior Analyst, Alternative Risk Premia · Machine learning and ESG  p.5
For more information please contact:
· What we offer  p.6
Clément Leturgie
Product Specialist, 1798 Alternatives
[email protected]
Machine learning for the prediction of stock returns

Introduction an uncommon scenario when excessive data mining leads to

models that work well on historical data, yet disappoint in
A mere decade ago, for most of us Machine learning (‘ML’) would
production.
be associated with pure science fiction. While ML methods have
been known for quite a while, it was only recently that the Machine learning can take the systematic approach to a whole
explosive growth of computational power accessible at low costs new level by minimising human participation at the stage of
facilitated the integration of ML in all aspects of our life. strategy development. Instead of manually searching for the best
Nowadays, numerous algorithms are running behind the scenes model, the researcher implements an ML algorithm that does this
filtering spam emails and enhancing our selfies. Armies of job in a more productive way. Of course, machines will never
developers have coded and optimised ML algorithms while succeed on their own. Even if all the tedious work is left to
academics provide plenty of educational material, making ML machines, the role of humans becomes critical in the design
accessible to a wide audience. of ML algorithms and data management. We will come back to
this later in the paper.
Media has further popularised ML and AI promising bright future
where machines perform all routine work, and even challenge Traditional fundamental analysis relies on simple manipulations,
humans in some creative activities. It is not surprising that ML with financial ratios to judge whether stocks are fairly valued or
has also become a very hot topic in finance. Over recent years not. In the systematic space, factor-based investing follows a
we witnessed an avalanche of research papers discussing the similar route by grouping financial ratios into buckets called
applications of ML to finance, including risk modelling, portfolio factors. Intuitively, it is difficult to believe that any simple predictive
construction, contextual analysis, and, of course the most relationships persist in the market and not arbitraged away. While
intriguing one, the prediction of asset returns. At the same time, we see no reason to question the long-term premia of factors,
we are aware of only a handful of successful implementations of systematic investors should be prepared to persevere through
advanced ML methods in actual portfolios. long periods of factors’ ups and downs before their patience is
rewarded. The recent disappointing performance of factor-based
At LOIM, we jumped on this train early on with the development
strategies is a timely reminder of this.
of a proprietary ML model (which we will refer to as ‘LOIM-ML’)
for stock scoring that we currently incorporate in our systematic Machine learning is here to help again. ML methods are designed
strategies. Truth be told, this was not an easy journey. During this to detect more complex predictive relationships, which are less
time, we experienced multiple frustrations and setbacks while likely to be easily arbitraged away, and therefore, are able to
discovering all the various pitfalls of implementation of ML on produce more stable returns. We expect that these new sources
financial data. In this note we would like to share our experience of alpha will provide a long-awaited diversification for systematic
and views. investment, which is almost fully dominated by factor-based
strategies. The flip side of the increased model sophistication is
Systematic investment without human biases the loss of its transparency. This is typically the main criticism of
Implementing quantitative models in asset portfolios is the ML as, traditionally, model interpretability is an integral part of
bread and butter of systematic investment. An example of such model validation. In our view, transparency is not critical for
an approach in equities is factor-based investing, which consists models “learnt” by machines since the automated process of
of building diversified portfolios with intentional exposures to model development minimizes human biases and judgement
certain systematic factors deemed to be rewarding. errors.1 That being said, it is possible to gain an insight into how
an ML model works through the analysis of its behavior.
In general, the systematic approach is a disciplined application
of a certain quantitative process. On paper, such an approach Machines will not work without human
avoids as much as possible human interventions thus ensuring
Advances in computer technology and the availability of open
the consistency of live performance with back-tested results.
source tools makes the application of Machine Learning seemingly
That said, even if the investment process is truly systematic, it has
straightforward. This is far from being true in general, and
been designed by humans. Efforts spent on the research model
especially in financial applications.
are often not very transparent and stay behind the scene. It is not

1
Interpretable Machine Learning is a new field that tries to address the problem of transparency where the model error can have severe consequences like, for example, self-driving cars.

For professional investor use only. Please see important information at the end of this document.
Page 2/8 Lombard Odier Investment Managers · Machine learning for the prediction of stock returns · May 2020
Machine learning for the prediction of stock returns

Financial theory teaches us that markets are efficient, therefore, went through a well-thought standardisation procedure making
asset returns are fundamentally difficult to predict. We should not predictors comparable across stocks and over time. Due to the
be overoptimistic to expect that the application of ML techniques dynamic nature of financial markets, magnitudes that seemed to
is going to change this dramatically. Instead, what we can hope be excessive in the past might be new normal today.
to achieve is a prediction accuracy that is good enough to produce
One of the biggest challenge of ML in finance is avoiding the
a decent performance when implemented in a diversified portfolio
forward-looking bias, which is a situation when the model learns to
of stocks. Good news is that only a tiny forecasting power is
predict the past from the future and not the other way round. The
sufficient. For example, a model predicting monthly stock returns
forward looking bias occurs when the data used for model training
with a hit ratio as low as 51% is enough to generate a Sharpe
was actually not known before the stock returns were realized. A
ratio of almost three2 ! The low forecasting power of financial
typical example is when a company quarterly earnings are used to
models is what makes the application of ML in finance especially
predict stock returns from the end of the same quarter onward. In
challenging where every decision counts.
real-life, quarterly figures are never available immediately. In fact,
Probably, the most important decision is the choice of ingredients this data may be even revised several times far in the future,
to be used in the model. A naïve approach to modelling stock which amplifies the forward looking bias when one uses the last
returns would be to throw all the possible stocks’ attributes into revision. Machine learning techniques are powerful enough to
the machine and let it do its magic. Unfortunately, this is not detect and take advantage of even slightest forward looking biases
going to work. First and foremost, we do not possess a sufficient resulting in overly optimistic back-tested results that cannot be
number of observations to deal with a large set of “raw” replicated in live. In LOIM, we avoid the forward-looking bias by
predictors. In image recognition, for example, we are able to taking into consideration the exact time when each quantity
estimate or train a model on an arbitrary large set of observations became publicly known or apply a conservative time lag when the
since experiments can be repeated as many times as needed. In exact date is unknown.
finance, we have to rely on historical observations only, which are
Machine learning algorithms are able to find arbitrary complex
naturally very limited. The dynamic nature of financial markets
relationships in data, which is both a benefit and a curse. If the
makes the reality even more complicated as we have to focus
learning process is not properly controlled, the algorithm may
on recent observations that are most relevant for the current
easily “overfit” data, which means discovering spurious
state of the markets.
relationships that are unlikely to hold in the future. This is even
In our view, the selection of model predictors should start by more important in finance where predictive relationships may
clearly formulating the hypothesis about the underlying model. In disappear due to ever changing markets. Exhibit 1 illustrates the
LOIM, we believe that stock returns are driven by fundamental and concept of overfitting using an example where an ML attempts to
technical forces, which is consistent with a long-standing culture find a relationship between a single predictor and an outcome. The
of stock picking. Inspired by a vast academic and practitioners’ “appropriate” model captures well the non-linear relationship in
literature as well as our internal expertise, we selected a total of data without becoming over sophisticated.
70 predictors of both technical and fundamental nature. All data

EXHIBIT 1 FINDING THE RIGHT BALANCE

UNDER-FIT APPROPRIATE OVER-FIT
Outcome

Outcome

Outcome

Predictor Predictor Predictor

Source: LOIM calculations, May 2020. For illustrative purposes only.

2
Source: LOIM calculations, based on the universe of stocks in the MSCI World index. For illustrative purposes only.

For professional investor use only. Please see important information at the end of this document.
Lombard Odier Investment Managers · Machine learning for the prediction of stock returns · May 2020 Page 3/8
Machine learning for the prediction of stock returns

The right level of model complexity is achieved by fine-tuning the by a tree diagram. In its simplest form, the decision tree has only two
learning process, which is called cross-validation. The idea behind possible outcomes depending on whether a certain condition is
this process is to split observations into two parts: training and satisfied or not. While such a tree is obviously too simple to make an
testing sets. The training set is used for model training, while accurate prediction, an aggregation of multiple trees like that can fit
testing set is left out for evaluating the model performance. This an arbitrary complex relationships in data.
data separation mimics what is happening in reality where the
Exhibit 2 illustrates how GBM works for a classification task
model is estimated on the past data while being implemented in
with two predictors. The process starts by training the first weak
the future. Therefore, the training set should strictly precede the
learner, which classifies observations based on a single criteria
testing set in time. The optimal configuration of the learning
involving values of Predictor 2. Next, the second weak learner is
process is such that it yields the best performing model on the
trained to explain the failure of the first one. The new classification
testing set and not the training one. The cross-validation is a very
model incorporates two criteria, one per each predictor. During
computer intensive process as it requires training multiples models
each subsequent learning cycle, a new weak learner is trained to
with different configurations of the learning process. At LOIM, we
improve or “boost” the accuracy of the model built so far. The
use cloud computing to perform cross-validation tasks.
process continues until the desired maximum number of learning
How do they learn? cycles is reached. In our example, after the third cycle, the model
with three criteria perfectly classifies the data.3
A good practice to introduce a complex subject such as Machine
learning is to start with a high level overview of different approaches What we have described above is, of course, an oversimplified
before going into details of particular algorithms. In this paper, we version of GBM. In real life applications, GBM receives many
decided not to do this as it is likely to be confusing for someone enhancement that help fine-tuning the learning process, which
who is not familiar with the subject. Instead, we directly introduce a can be a game changer especially in financial applications.
particular methodology called Gradient Boosting Machines (GBM). Ultimately, the design of the ML algorithm and the cross-validation
procedure are those human decisions that determine the success
Despite its somewhat confusing name, Gradient Boosting Machines
of machines. On our side, we did a comprehensive analysis of a
is based on an intuitively straightforward idea of building a complex
number of different Machine Learning models and selected GBM
model by aggregating multiple simple ones. In most applications,
for the LOIM-ML predictor model. This is also in line with GBM
these simple models, also called weak learners, take the form of a
being one of the top choices among practitioners.4
decision tree – a model whose prediction can be visually represented

EXHIBIT 2 GBM: FROM WEAK LEARNERS TO A STRONG MODEL

Predictor 2

Predictor 2

Predictor 2

Predictor 1 Predictor 1 Predictor 1

Source: LOIM calculations, May 2020. For illustrative purposes only.

3
Any further learning will improve the classification in terms of the likelihood of the prediction as the outcome of the model is the probability of each class.
4
A forthcoming technical paper will provide a more comprehensive analysis of different alternatives.

For professional investor use only. Please see important information at the end of this document.
Page 4/8 Lombard Odier Investment Managers · Machine learning for the prediction of stock returns · May 2020
Machine learning for the prediction of stock returns

LOIM-ML model: new alpha or old factors? Economic, Social and Governance (ESG) is probably the most
distinguished data providing an assessment of various ESG issues
Our LOIM-ML model makes predictions based fundamental
in the form of quantitative scores.
and technical stock characteristics, which overlap with definitions
of traditional equity factors. One may wonder whether the model The key distinction of ESG data from traditional fundamental and
discovers new sources of alpha or it simply rediscovers the technical characteristics is that it is a qualitative in nature, and
“old” factors. To answer this question Exhibit 3 shows the full therefore tends to be sensitive to the scoring methodology. For
performance of a long-short portfolio built using the LOIM-ML example, it is not uncommon that different ESG vendors disagree
model score,5 and the part explained by time-varying portfolio substantially in their assessments of ESG issues. In our view, ESG
exposures to five major factors: Value, Quality, Momentum, Low data provides a rich alternative set of stock characteristics that
beta and Small size.6 can be very useful for return prediction. However, a short history
of the data and its qualitative nature prohibits implementing
Knowing that factors have performed well in the past, it would
complex predictive models.
be surprising if machines did not learn their predictive abilities.
Indeed, we observe that the exposure to factors explains about one At LOIM, we built a proprietary ESG-QUANT score using
third of the total return of the strategy since 2004. However, it is elements of Machine learning. Our ESG-QUANT score represents
also clear that from 2016 onward, factors do not add any value an index of original ESG scores with weights reflecting their
while the “new alpha” continues to deliver a robust performance. capacity to predict stock returns. We require that each ESG score
Overall, LOIM-ML model meets our expectations that Machine has a positive or no contribution to the aggregate score, which is
learning is able to discover more complex predictive relationships a reflection our conviction that companies with fewer ESG-related
beyond those used in simplistic factor-based models. issues are likely to outperform. The methodology is described in
more details in our recent white paper “ESG Alpha: Doing Well
Machine learning and ESG While Doing Good”.
Recently, we have witnessed the emergence of multiple alternative
data sources that offer new insights on the business of companies.

EXHIBIT 3 LOIM-ML MODEL LOOKS BEYOND FACTORS

60%
Strategy Factors New alpha
50%
40%
30%
20%
10%
0%
-10%
2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

2018

2019

Source: LOIM-ML. For illustrative purposes only.

5
This is an unconstrained long-short portfolio where stock weights are proportionate to their scores from LOIM-ML model.
6
We used proprietary definitions.

For professional investor use only. Please see important information at the end of this document.
Lombard Odier Investment Managers · Machine learning for the prediction of stock returns · May 2020 Page 5/8
Machine learning for the prediction of stock returns

What we offer Our research efforts in Machine learning did bear fruit. We are
pleased to announce a recent launch of a new investable long-
In our everyday life, we are witnessing an increasing adoption of
short equity strategy as a building block of the ARP offering.
data-intensive technologies and a rising interest in data science.
This strategy seeks to achieve a dual objective of “doing well”
Finance is not an exception. Machine learning opens new horizons
(generating performance) and “doing good” (promoting responsible
for systematic investment, however its successful application requires
companies) by exploring opportunities both on the long and on
much more effort than just hiring a bunch of data science graduates.
the short side. The application of ML techniques to ESG data
In this paper, we highlighted the main promises and challenges
(ESG-QUANT score) establishes the essential balance between the
of using ML techniques for prediction of asset returns. The main
two objectives. ML approach to fundamental and technical data
message we wanted to deliver here is that only a combination of
further enhances the strategy “doing well” side (LOIM-ML model).
strong financial expertise with a good knowledge of the data science
will produce a strategy that performs not only on the paper. The success of any systematic strategy depends just as much
on the quality of the model as on the efficiency of portfolio
construction and execution. A model that looks attractive on the
EXHIBIT 4 LOIM-ML EQUITY ESG ALPHA STRATEGY paper will not always lead to a decent strategy as the alpha gets
INNOVATIVE ALPHA ENGINE ESG PILLAR INTEGRATION quickly “eaten” by multiple layers of deductions such as direct and
indirect transactions costs, dividend taxes, shorting fees and
ML Equity ESG-CAR operational costs. In LOIM we optimize the strategy execution using
an advanced procedure that allows achieving given performance
Controversies and impact objectives in a most diversified and cost efficient way.
ML ESG Non sustainable
business
Exhibit 4 presents an overview of the main building blocks of
the strategy. A forthcoming paper will review the implementation
Execution Carbon emission process from the model development to portfolio construction and
execution. In the meantime, we will be happy to provide any details
and explanations on request.
RESPONSIBLE UNCORRELATED RETURNS

Source: LOIM. For illustrative purposes only.

For professional investor use only. Please see important information at the end of this document.
Page 6/8 Lombard Odier Investment Managers · Machine learning for the prediction of stock returns · May 2020
Machine learning for the prediction of stock returns

Important information
For institutional investors only. Neither this document nor any copy thereof may be sent, taken into, or distributed in
This document has been issued by Lombard Odier Funds (Europe) S.A. a Luxembourg the United States of America, any of its territories or possessions or areas subject to its
based public limited company (SA), having its registered office at 291, route d’Arlon, jurisdiction, or to or for the benefit of a United States Person. For this purpose, the term
1150 Luxembourg, authorised and regulated by the CSSF as a Management Company “United States Person” shall mean any citizen, national or resident of the United States
within the meaning of EU Directive 2009/65/EC, as amended; and within the meaning of America, partnership organized or existing in any state, territory or possession of the
of the EU Directive 2011/61/EU on Alternative Investment Fund Managers (AIFMD). United States of America, a corporation organized under the laws of the United States
The purpose of the Management Company is the creation, promotion, administration, or of any state, territory or possession thereof, or any estate or trust that is subject to
management and the marketing of Luxembourg and foreign UCITS, alternative United States Federal income tax regardless of the source of its income.
investment funds (“AIFs”) and other regulated funds, collective investment vehicles Source of the figures: Unless otherwise stated, figures are prepared by LOIM.
or other investment vehicles, as well as the offering of portfolio management and Expected returns are estimates of hypothetical average returns of economic asset
investment advisory services. classes derived from statistical models. There can be no assurance that these returns
Lombard Odier Investment Managers (“LOIM”) is a trade name. can be achieved. Actual returns are likely to vary.
This document is provided for information purposes only and does not constitute an Models may be misspecified, badly implemented or may become inoperative when
offer or a recommendation to purchase or sell any security or service. It is not intended significant changes take place in the financial markets or in the organization. Such
for distribution, publication, or use in any jurisdiction where such distribution, a model could unduly influence portfolio management and expose to losses.
publication, or use would be unlawful. This material does not contain personalized Although certain information has been obtained from public sources believed to be
recommendations or advice and is not intended to substitute any professional advice on reliable, without independent verification, we cannot guarantee its accuracy or the
investment in financial products. Before entering into any transaction, an investor should completeness of all information available from public sources.
consider carefully the suitability of a transaction to his/her particular circumstances Views and opinions expressed are for informational purposes only and do not constitute
and, where necessary, obtain independent professional advice in respect of risks, as well a recommendation by LOIM to buy, sell or hold any security. Views and opinions are
as any legal, regulatory, credit, tax, and accounting consequences. This document is the current as of the date of this presentation and may be subject to change. They should
property of LOIM and is addressed to its recipient exclusively for their personal use. not be construed as investment advice.
It may not be reproduced (in whole or in part), transmitted, modified, or used for any No part of this material may be (i) copied, photocopied or duplicated in any form, by
other purpose without the prior written permission of LOIM. This material contains the any means, or (ii) distributed to any person that is not an employee, officer, director, or
opinions of LOIM, as at the date of issue. authorised agent of the recipient, without Lombard Odier Funds (Europe) S.A prior consent.
©2020 Lombard Odier IM. All rights reserved.

For professional investor use only. Please see important information at the end of this document.
Lombard Odier Investment Managers · Machine learning for the prediction of stock returns · May 2020 Page 7/8
www.loim.com