Improving the prediction of asset returns with machine learning by using a

custom loss function

Jean Dessain
IESEG, School of Management, Department of Finance, 3 rue de la Digue, 59000 Lille, France
Email: [email protected]

Abstract
Not all errors from models predicting asset returns are equal in terms of impact on the
efficiency of the algorithm: some errors induce poor investment decisions while other errors
have no financial consequences. This asymmetry is critical for the performance metric used to
assess the ability of algorithms to predict returns. This should also impact the choice for the
most efficient loss function, a key element for training machine learning algorithms.
Mean Squared Error (MSE) is the most popular loss function for regression algorithms, but
it is not the most efficient one for algorithms predicting asset returns. In this article: (a) we
develop custom loss functions that account for the asymmetry in the predictive purpose of the
algorithm. (b) We compare the efficiency of these custom loss functions with MSE. (c) We
present an efficient custom loss function that significantly improves the prediction of asset
returns, and which we confirm to be robust.

Keywords: Time series forecasting; Stock return predictability; Investment efficiency; Machine
learning; Deep learning; Loss function
JEL: C45, C53, G11, G17, N2.

1. Introduction
The finance industry has systematically looked for ways to predict future asset returns, and
more generally to predict financial time series data. Regression and classification algorithms,
as well as reinforcement learning, have been developed to predict the effective return of assets,
either for very short periods of time or for longer horizons. But the task is undoubtedly difficult
as financial markets are volatile and noisy environments, with short-term and long-term
fluctuations and huge shifts in volatilities.

Electronic copy available at: https://ssrn.com/abstract=3973086

1.1. Situation
Numerous academic papers present algorithms aimed at improving investment strategies
and optimizing the risk-adjusted returns of portfolios. The number of academic studies
published on this topic has grown at an exponential rate and a comprehensive review of the
literature is becoming increasingly challenging (Bustos and Pomares-Quimbaya, 2020; Huang
et al., 2020; Meng and Khushi, 2019; Ozbayoglu et al., 2020; Sezer et al., 2020; Thakkar and
Chaudhari, 2021), if feasible at all.
(Dessain, 2022) offers arguably the most comprehensive overview to date, with 190 articles
reviewed over the period 2010 – June 2021, but with a narrow focus on the sole performance
metrics used to compare algorithms predicting asset returns. He demonstrates that the mean
squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE) or similar
error-based performance metrics are not appropriate for assessing the results of algorithms
predicting asset returns. The underlying reasoning is that error-based metrics treat all errors
equally and do not differentiate an error that triggers a bad investment decision (an investment
resulting in a negative return or a missed opportunity with no investment when the asset has
led to a positive return) from a prediction error which does not have any adverse consequence,
leading to a positive return or to a non-investment that avoided a negative return. The deductive
reasoning is confirmed with an empirical analysis.
The same MSE is the most common loss function in regression algorithms. A loss function
measures the prediction error of the model that the optimization process tries to minimise in
updating the weights during the backward propagation. If MSE is inappropriate to assess the
efficiency of an algorithm, our assumption is that MSE might not be the optimal loss-function
for training a neural network performing a regression to predict asset returns. Indeed, as for the
performance metric, all errors computed with MSE (and RMSE, MAE or MAPE) will be
treated equally during the training phase, even though their consequences might be drastically
different in terms of investment efficiency.

1.2. Contribution
This paper analyses how a custom loss function can be organised to render the asymmetry
of the objective function and can be implemented in deep neural networks.
The rest of the paper is structured as follows: In section two, we analyse the literature review
in terms of custom loss functions for deep learning, with models predicting asset returns as well
as models from other fields of computer science. Section three presents the framework and
various custom loss functions as alternatives to the MSE. Section four presents the
methodology for testing the various loss functions. Section five presents the results of the
analyses and demonstrates that a custom loss function significatively overperforms MSE.
Section six concludes and draws some perspectives.

Electronic copy available at: https://ssrn.com/abstract=3973086

2. Common and custom loss functions in the literature
We analyse the loss functions in the literature from four sources: (i) articles analysing the
loss functions in prediction problems without referring to machine learning algorithms, (ii) the
database of 190 articles predicting stock returns used by Dessain (2022), (iii) some articles of
machine learning in finance not covered in (ii) and that specifically analyse custom loss
functions, and (iv) articles from other fields of AI that explicitly refer to the use of custom loss
functions. Besides (ii), we search with Google Scholar for the most relevant articles with
“custom loss function”, “asymmetric loss functions”, and performing the same search replacing
loss with cost.

2.1. Asymmetric losses besides ML models

The concept of asymmetric loss is not new. Zellner (1986) tests an asymmetric loss function
called Linear Exponential, often referred to as “LinEx” loss. This loss function is a key element
in Patton and Timmermann (2007) analysis of the properties of optimal forecasts under
asymmetric loss and non-linearity. They demonstrate the properties of the loss, including its
differentiability. Among other time-series, Patton and Timmermann test the LinEx loss
function versus the MSE for various applications, including for predicting the annualised return
of Exxon stock over a period from 1970 to 2003.
LinEx loss has been widely tested since Patton and Timmermann. Christoffersen and
Diebold (1996) use LinEx for solving prediction problems of asset returns. Hwang et al. (2001)
extend the results to forecasting a volatility process as the sum of conditional volatility and an
adjustment factor.
Singh (2014) attempts a comprehensive overview of loss functions used in the financial
sector. Besides the symmetric MSE, MAE and all their variations, he noticed three asymmetric
losses: the LinEx, Linear-Linear and Quadratic-Quadratic losses. Singh presents the LinEx as
“particularly well-suited in financial applications”.
Nevertheless, neither the LinEx nor the Linear-Linear and Quadratic-Quadratic loss
functions do address our purpose: in each case, the asymmetry is linked to the sign of the loss
(y - 𝑦). The asymmetry of our problem is linked to the difference in sign between the actual
return y and its predicted value 𝑦, or to the sign of the product (y * 𝑦). The sign of the loss is
irrelevant for the economic efficiency of the algorithm.

2.2. Financial articles related to algorithms predicting stock returns

The loss function is, most of the time, not considered as a relevant topic. Using the
aforementioned database, we examined each of the 190 articles and searched for the loss
function(s) applied. 39 articles1 describe or mention the loss function(s) applied, and 151 do
not address the topic at all. 20 out of the 39 articles perform regressions (sometimes together

1
See Appendix 1. Loss functions applied by author. For those who do not describe the custom loss function,
they are out of our scope and we simply refer to the bibliography of Dessain (2022)

Electronic copy available at: https://ssrn.com/abstract=3973086

with classification tasks). Out of these 20 articles, 13 apply standard MSE or RMSE. 3 articles
apply MAE or MAPE.
Four articles directly or indirectly investigate a possible customisation, but none of them
introduce asymmetry in the loss function.
Gu et al. (2020) test various loss functions, starting with MSE, then using weighted MSE to
underweight small stocks in favour of large stocks. They also test a Huber loss function as a
mix of MSE (for the relatively small errors) and MAE (for relatively large errors). All errors
are treated equally, without any consideration for the sign of the expected return versus the
actual one.
Lim et al. (2019) enrich the MSE loss with a regularisation based on the Sharpe ratio. All
errors are treated equally, without any consideration for the sign of the expected return versus
the actual one. They obtain interesting results on 5 years out-of-sample data, as long as
transactions costs are at 0, but they obtain negative Sharpe ratios for each neural network
architecture (MLP, CNN or LSTM)2 as soon as transaction costs reach or exceed 5 basis points.
Yun et al. (2020) combine the standard MSE to a directional cost in view of managing a
portfolio of assets via a two-stage deep learning approach. For each asset within the portfolio,
the joint cost function considers its absolute and relative performances, but it does not consider
any asymmetry in the impact from the prediction error.
Zhou et al. (2018) combine a forecast error loss, MSE with a directional loss in a GAN3
architecture that includes LSTM and CNN layers for high-frequency data. To the forecast error
loss, they add a loss that tracks a directional change. The performance metrics applied by Zhou
et al. (2018) are the root mean squared relative error (RMSRE) and the direction prediction
accuracy (DPA), an error-based metric and an accuracy-based one, whose inefficiency have
been demonstrated by Dessain. As the metrics are unreliable, the true merit of the adjustment
proposed by Zhou et al. can therefore not be confirmed.

2.3. Other financial articles

We also reviewed custom loss functions in articles dedicated to finance but not covered in
2.2. We found three papers that investigate custom loss functions.
Dress et al. (2018) test various asymmetric loss functions with linear regressions, decision
trees, SVM and MLP to forecast the value of resold leasing cars. They compare the LinEx loss
with a Quadratic-Quadratic Loss “QQL” function. As for LinEx, the trigger for asymmetry of
the QQL is determined by the sign of the error, making QQL inadequate for our purpose.
Ahmed et al. (2020) refer to the fact that algorithms and their constituents should be tailor
made in respect of the field of application. They apply a custom loss function called Forex Loss
Function (FLF) to predict forex movements with a LSTM model. Instead of computing the loss
with the difference between the predicted exchange rate and the actual one, they adjust the MSE

2
MLP = Multi-Layer Perceptron, CNN = Convolutional Neural Network (in this case wavenet) and LSTM =
Long-Short Term Memory neural network, a form of Recurrent Neural Network (RNN).
3
GAN = Generative Adversarial Network, here with an LSTM generator and a CNN discriminator.

Electronic copy available at: https://ssrn.com/abstract=3973086

using the Open – High – Low – Close (OHLC) predicted prices and OHLC actual ones (inspired
by the candlestick analysis). They then use MAE as performance metric to assess the higher
efficiency of the custom FLF, leading to non-reliable results.
Tavakoli and Doosti (2021) propose a custom loss function that adjusts the MSE to account
for (i) the expected future volatility of the asset and (ii) the expected gain defined a priori as
target for the model. They compare the accuracy of the model with various standard losses
(MSE, MAE, MAPE, Huber loss, Log-Cosh loss). They find a higher accuracy for the proposed
loss function and assume that the superior accuracy should lead to a higher profit. The test set
(out-of-time) is three months of weekly returns or 13 data points. The approach is interesting
but should be further confirmed, as 3-months test set of weekly data is obviously too short to
draw a reliable conclusion, and the chosen performance metric cannot lead to any effective
conclusion.

2.4. Other non-financial articles with custom loss functions

Eventually, we reviewed custom loss functions in other areas of research. From our
analysis, medical and biomedical field and time series predictions are the main fields to propose
several approaches, even if the loss function is not a popular theme.
Barton et al. (2021) use a weighted combination of global and local MSE to improve signal
detection in biomedical science, using the combination of two symmetric loss functions to
induce asymmetry. Bhandari et al. (2020) combine a regression and a classification approach,
with an MSE and a cross-entropy loss function to improve the detection of oral tumours. Raein
Hashemi et al. (2019) improve both the classification of lesions and the detection of multiple
sclerosis using focal loss with MLP and CNN algorithms.
To predict time-series of various nature (astronomy, geology, medical imaging), Cuturi and
Blondel (2017) apply a soft dynamic time-wrapping (soft DTW) that allows to compare time
series of variable size with a differentiable loss. The exercise is purely academic, and the times-
series serve as a test with no other purpose.
Interestingly, Chen et al. (2021) explicitly search for a loss function tailor-made for wind
speed prediction. A kernel-MSE is applied to cope with the nonlinearity of wind speed data
forecasting. The kernel MSE is symmetric, positive, bounded, with a positive first order
derivative and a negative second order derivative; therefore, it should be less sensitive to
outliers than the standard MSE.
Eramo et al. (2021) propose an asymmetric loss function to predict traffic and cloud
resource requirements for allocating cloud and bandwidth resources. The asymmetric loss
function captures the dramatic increase in costs when the prediction error leads to the under-
capacity of the cloud and to an excess traffic compared to the available bandwidth. The
proposed model with asymmetric loss function has led to a significant cost advantage compared
to standard models. Wu et al. (2021) applies an asymmetric “Linear-Linear Loss” (LLL)
function to a support vector regression to optimize the electric load forecasting. As with LinEx,
LLL asymmetry triggers where the loss changes of sign.
Finally, Ebert-Uphoff et al. (2021) propose a guide and a large collection of loss functions
applied in environmental science, stating that “the loss functions required by environmental

Electronic copy available at: https://ssrn.com/abstract=3973086

scientists are unlike any loss functions typically used in computer science, and the community
has not yet developed comprehensive resources, such as a large collection of customised loss
functions.”
While none of these approaches are immediately relevant for our task of predicting returns,
the process of tailoring the loss function for a specific purpose echoes our initial intuition.

3. Framework and custom loss functions

The prediction of asset return (or asset price) is driven by the asymmetry of goals. If the
prediction error induces a wrong investment decision, its impact is significant and leads either
to a loss-making investment or to a missed profit-making investment. If the error does not result
in a wrong investment decision, its impact is negligible. Therefore, the optimization should not
minimise the mean error, but only the mean of the errors that cause a wrong investment
decision. To achieve this objective, the investment strategy should be defined ex-ante, before
the tailoring of the loss function.

3.1. Framework: data and investment strategy

To test our hypothesis of superior efficiency of a custom loss function, we define a
framework. We will test various loss functions on a series of stocks whose daily returns are
predicted by machine learning regression algorithms.
We use a long series of daily prices history to train our algorithm, and we will test4 it with
out-of-time data that will be long enough to ensure that most typical situations are properly
represented: rally, crisis, recovery, periods of uncertainty with no clear direction, high and
low volatility periods. This is a necessary condition to produce reliable results 5.
In both training and testing phases, the algorithm will receive a set of historical data up to
the end-of-day Closet and will predict the next day's return. A transaction cost is charged for
any purchase or sale of shares: if an open position exists and is rolled over, the transaction cost
does not apply; if the position is either opened or closed, the transaction cost is charged.
We test a very simple investment strategy: (i) if the predicted return of the next day is above
a defined trigger, we invest or roll the open position, if any, for one day; (ii) otherwise, we take
no open position or we close any previous open position.

4
We do not use the traditional split between training set, validation set, and test set as we are just looking for
algorithms to produce series of returns.
5
This is a recurrent issue with the 190 surveyed articles in 2.1, where many papers apply a test set insufficiently
long, sometimes even equal or shorter than 2 years, thereby producing unreliable and not exploitable results.

Electronic copy available at: https://ssrn.com/abstract=3973086

3.2. Analysed loss functions
To match the defined investment policy, the custom loss function should penalise errors
when the effective daily return y and the predicted return 𝑦 are of opposite signs, and
underweight loss when both actual and predicted returns are of the same sign.
We compare our analysis with the MSE loss function and design 3 main types of
customisation:
(i) an “adjusted loss 1” (AdjLoss1) where the squared error is multiplied by α if (y * 𝑦) < 0
(when y and 𝑦 are of opposite sign), and by 1 / α if (y * 𝑦) ≥ 0 (when y and 𝑦 have the same
sign), with α being a positive number greater than 1. (α =1 is equal to MSE). This AdjLoss1 is
not fully differentiable, where either y or 𝑦 is equal to 0.
(ii) an “adjusted loss 2” (AdjLoss2) where the squared error is multiplied by α if (y * 𝑦) < 0
(when y and 𝑦 are of opposite sign), and by 1 if (y * 𝑦) ≥ 0 (when y and 𝑦 have the same sign),
with α being a positive number greater than 1. This AdjLoss2 is a variant of AdjLoss1 as the
loss is modified only when y and 𝑦 do not have the same sign.
(iii) an “adjusted loss 3” (AdjLoss3) that, to some extent, mimics the ReLu activation
function where the squared error is multiply by (1+γ) if (y * 𝑦) < 0 (when y and 𝑦 are of
opposite sign) and multiplied by γ if (y * 𝑦) > 0 (when y and 𝑦 have the same sign), with γ
being a positive number smaller than 1. This AdjLoss3 is not fully differentiable, when either
y or 𝑦 is equal to 0.
We also designed an “adjusted loss 4” (AdjLoss4) that is fully differentiable. The squared
error is multiplied by a sigmoid-like curve adjustment factor. This adjustment factor tends to a
maximum for large negative values of (y * 𝑦), is equal to the 1 for y or 𝑦 equal to 0, and tends
towards a minimum for large positive values of (y * 𝑦).
∗( )
𝐴𝑑𝑗𝐿𝑜𝑠𝑠4 = ( . ) with 𝜕 being a positive number greater than 1.
( )
( ( ∗ ∗ )

Eventually, we tested the LinEx loss function made popular by Patton and Timmermann
(2007) : LinEx = 𝑒 ∗( )
− 𝑎 ∗ ( 𝑦 − 𝑦) − 1. We obtained irrelevant results for the
purpose of our analysis.

Visualisation of the loss functions

We compute the error values for 6 loss functions: MSE, AdjLoss1 with α = 2 and α = 1.5 ,
AdjLoss2 with β = 2.5 and β = 2.25 and AdjLoss3 with γ = 0.1.
For the sake of completeness, we also present the two fully differentiable functions AdjLoss4
with 𝜕 = 4 and LinEx with a = 2.
The curves of the various custom loss functions are presented in Graph 1. and Graph 2. MSE
is the only symmetric loss function. Compared to MSE, AdjLoss1, AdjLoss2 and AdjLoss3
increase the loss when signs of actual and predicted returns are opposite and reduce the loss
when actual and predicted returns have same sign. AdjLoss4 and Linex increase the loss in both
cases in an asymmetric manner. The non-complete differentiability of AdjLoss1, AdjLoss2 and
AdjLoss3 is visible with Graphs 1 and 2.

Electronic copy available at: https://ssrn.com/abstract=3973086

Graph 1. Loss level per loss function with a true return of 1.00%

Graph 2. Loss level per loss function with a true return of -1.00%

Electronic copy available at: https://ssrn.com/abstract=3973086

4. Analysis
To compare the efficiency of the various loss functions, we organise our process as described
hereafter, with the data management, the DL model and the performance measurement.

4.1. Data collection and data management

We collect publicly available data for large stocks traded on NYSE and Euronext from
01/01/1996 to 31/12/2020, and we reject any stock whose historical series is shorter than 15
years. We end up with a list of 105 stocks6: 43 traded in USD on NYSE and 62 traded in EUR
on Euronext. As explained in 3.1, we organise for each stock a X matrix that is made of (i)
OHLC+V daily data, (ii) a set of 14 technical indicators, and (iii) the closing prices of all the
other stocks traded on the same exchange. In total, X contains 61 inputs for US stocks and 80
inputs for EU stocks. The array Y is made of the daily returns of the stock between the next
opening and the opening of the day after: 𝑦 = − 1. The choice of considering the
return from the next opening Opent+1 until the opening of the day after Opent+2 adequately
integrates operational constraints such as the time required for collecting all the data and
running the model, gathering all investment decisions and having controls performed before
the execution of the generated buy or sell orders.
X and Y are split between an “in-time” train set (X_train and y_train) and an “out-of-time”
test set (X_test and y_test. The size of the out-of-time test set for each stock is 1260 trading
days, or five years of trading, between 01/01/2016 and 31/12/2020.
Graph 3. shows the cumulative result as percentage of the invested amount for buy & hold
portfolios equally invested in the 43 US stocks and, respectively, for a portfolio equally
invested in the 62 European stocks, between 01/01/2016 and 31/12/2020. The test set includes
the various types of market sentiments we want to have to ensure the reliability of our results,
including a severe crisis and a rebound, high and low volatility periods, etc.

6
See Appendix 2. List of stocks. The data set has been published on Mendeley Data doi:10.17632/nbwhzctrjp.2

Electronic copy available at: https://ssrn.com/abstract=3973086

Graph 3. Cumulative results of buy & hold portfolios over 5 years

Min-max normalisation is applied to X_train. The min and max values derived from X_train
are applied to X_test for the min-max normalisation of the test set.

4.2. Model
In order to test the various loss functions, we tested various straightforward MLP models,
with all interconnected neurons, and one single neuron as output layer for the regression. We
present herewith the results of one architecture with 8 hidden layers. Other tested
architectures, from 2 to 16 hidden layers, delivered very similar results in term of relative
efficiency of the various loss functions.

10

Electronic copy available at: https://ssrn.com/abstract=3973086

Graph 4. Simplified structure of the MLP for testing the loss functions.

The activation function is the standard Rectified Linear Unit (ReLu). With 61 inputs for
US stocks and 8 hidden layers, we obtain 188.893 learnable parameters. The number of
parameters increases to 193.776 with the 80 inputs for EU stocks.

Table 1. Number of learnable parameters of the models

We run the model several times on several computers with various hyper-parameters to
verify:

11

Electronic copy available at: https://ssrn.com/abstract=3973086

 - The reproducibility7 of the model: that secures exact same results when ran over the
same computer with the same parameters;
- The replicability of the model: that leads to very similar results when ran over different
computers and/or with marginally different hyper-parameters 8.
This way, we make sure that the results are reliable and consistent over time.

4.3. Investment strategy

To benchmark any strategy with the buy & hold strategy, for each stock, we repeat this
investment process: a “reference number of shares” is initially determined as the number of
shares that can be purchased with 100.000 USD (or 100.000 EUR for the European shares) on
01/01/2016.
This reference number of shares is kept as an open position over the 5 years until
31/12/2020 for the buy & hold strategy. The result of the buy & hold strategy will serve as
common benchmark to compare the relative efficiency of the algorithms.
For each day of the out-of-time test set, if the return predicted by the ML algorithm is
above -0.20% (that is 2 times the transactions costs of 0.10%), the algorithm invests in the
asset: either it rolls over at no cost an existing open position or it purchases the shares with a
transaction fee of 0.10% on the value of the deal. The number of shares purchased is the
lower of (i) the reference number of shares and (ii) the maximum number of shares that could
be purchased with the outstanding cash position for the stock resulting from previous
investment decisions.
If the predicted return is smaller than -0.20%, the algorithm either sells the entire existing
open position with a transaction fee of 0.10% on the value of the sale or it maintains at no
cost the open cash position.
The 0.10% fee is above what is typically paid by small institutional investors. It
incorporates the direct costs paid to brokers and the indirect costs that might be linked to the
lower liquidity at opening. Cash that is not invested in the stock is kept available without any
remuneration.

4.4. Performance criteria

The objective is to assess whether each custom loss function does significantly perform
better than the standard MSE loss function, on average and consistently across the 105 stocks.
For each of the 6 loss functions and for each of the 105 stocks9, we compute the daily returns

7
In order to get reproducibility and replicability, we fix the seed for Python, for Numpy and for Pytorch.We also
force Pytorch to work with deterministic CUDNN. Details about hardware and software are provided in
appendix
8
We tested the model with various learning rates between 0.0001 and 0.001, with dropout values between 0.15
and 0.30, and with epochs between 100 and 300. We vary the number of hidden layers between 10 and 6.
9
We therefore compute 630 series of 1260 daily returns generated by algorithms, and 105 daily returns for the
buy & hold strategy.

12

Electronic copy available at: https://ssrn.com/abstract=3973086

realised on the stock and we compare them with the returns achieved with the buy & hold
strategy, or 105 series of 1260 daily returns per loss function.
To test the effectiveness of each tested loss functions, we compare (i) the risk-adjusted
performance of the average return generated by the AI algorithm with (ii) the risk-adjusted
average return of the buy & hold strategy. Rather than using a Sharpe ratio that assumes the
normality of the returns’ distribution, we prefer to assess the riskiness of the performance with
the Value-at-Risk adjusted with the Cornish-Fisher expansion (CF-VaR)10. This way, we
properly account for the effective skewness and kurtosis of the distribution of asset returns.
This relative performance metric, called D ratio, appears to be among the best performing
metric to assess machine learning algorithms predicting asset returns (Dessain, 2022), as it
provides a metric that measures the true added value of an AI algorithm compared with a neutral
reference (the buy & hold strategy), independently from the market environment. The D ratio
indicates the extent by which the risk-adjusted return of the algorithm exceeds the risk-adjusted
return of the buy & hold strategy: if the D ratio is above 1, the algorithm outperforms the buy
& hold strategy; if it is below 1, the algorithm underperforms the buy & hold.
To complement the main metric, we also compute two sub-metrics: (i) the D-Return ratio
which measures the average return generated by the algorithm divided by the average return
of the buy & hold strategy, and (ii) the D-VaR ratio that measures the relative risk (CF-VaR)
of the returns from the buy & hold divided by the CF-VaR of the returns of the algorithm.
With each sub-metric, a result above 1 indicates that the algorithm overperforms the buy &
hold strategy, respectively with higher returns or with lower riskiness.
To test the stability over time and the reliability of the performance of the algorithm with
each loss function, each metric (D ratio, D-Return and D-VaR) is computed on the entire 5
years period, and on two sub-periods of 2.5 years.
Once the D ratio, D-Return and D-VaR have been computed for each stock and each
analysed loss function, we measure the robustness of each loss function and its consistency:
for each loss function, we rank the 105 D ratios (respectively D-Return and D-VaR) and build
a graph with the cumulative distribution of D ratios (respectively D-Return and D-VaR) for
various values of D ratios ranking from -5 to +5. We then compute the area-under-curve
(AUC) for each cumulative distribution of D-ratios. The best loss function is the loss function
with the highest D ratios and therefore the smallest AUC. The smallest AUC indicates the
highest frequency of high D ratio values (respectively D-Return and D-VaR) and therefore
the best performance algorithm. The highest AUC translates the highest number of lowest D
ratios (respectively D-Return and D-VaR). This way, we can assess the robustness of the loss
function across the 105 stocks. A loss function with a very low AUC performs better than a
loss function with significantly higher AUC. To make the comparison easier, we normalise
the AUC to 1 for the buy & hold strategy. An AUC below 1 indicates that the model

10
D ratio follows the same principle as the Sharpe ratio, but the risk measure (CF-VaR) is more robust than the
standard deviation applied by Sharpe, as it does not assume a normal distribution of returns, an assumption that is
most of the time not verified in practice.

13

Electronic copy available at: https://ssrn.com/abstract=3973086

outperforms the buy & hold strategy; an AUC above 1 shows that the model underperforms
the buy & hold strategy.

5. Detailed empirical results

The objective is to compare the efficiency of the various custom loss functions to the
standard MSE, and to assess the impact of the loss functions on the performance of the
algorithm. We first analyse whether the results of the algorithm with different custom loss
functions are statistically different from the MSE. We then check whether the average
performance of the loss functions is above the performance of the MSE loss. We eventually
verify that the superiority is consistent across most of the individual stocks, and we compare
the distribution function of the performance metric of each loss function.

5.1. Dissimilarity of the D ratio from custom functions

To test the statistical difference between the series of 105 D ratios obtained for each loss
function, we use a Mann-Whitney U test (MWU test) which does not require any assumption
about the implicit distribution function of the series. MWU test is applied for each series
against the D ratios obtained with the MSE loss function.
We obtain p-values disclosed in Table 2. that are below a significance level of 0.01 for all
loss functions, except for AdjLoss2 with β equal to 2.25, that is only significantly different
from MSE at a confidence level of 0.065. The results with AdjLoss4 were not reliable and
have been rejected from the analysis.
Table 2. p-values of the Mann-Whitney U test
p value of
Loss function
MWU test
AdjLoss1 2 8.54E-06
AdjLoss1 1.5 2.01E-06
AdjLoss2 2.5 4.19E-07
AdjLoss2 2.25 0.065475
AdjLoss3 0.1 1.90E-05

We conclude that the D ratios obtained with the various custom loss functions are
statistically different from the D ratios obtained when the algorithm runs with the MSE loss
function.

5.2. Average D ratio

The best algorithm has the highest and most stable D ratio, that is the highest expected
return adjusted for risk relative to the risk-adjusted return of the buy & hold strategy.
As shown in Table 3., with an average D ratio of 0.217, the MSE loss function achieves
the lowest average D ratio of all tested loss functions. Every custom loss function achieves
better results than MSE. Not only is it true for the entire period, it is also true with most

14

Electronic copy available at: https://ssrn.com/abstract=3973086

custom loss functions for each sub-period. This confirms our intuition that asymmetric loss
functions over perform the MSE.
The best loss function is the AdjLoss2 with β equal to 2.5 that achieves an average D ratio
of 1.585 over the 105 stocks. With 1.081 for the first sub-period and 1.819 for the second sub-
period, the AdjLoss2 with β of 2.5 overperforms the MSE for each sub-period. The D ratio
above 1 also indicates an average risk-adjusted performance superior to the buy & hold
strategy.
AdjLoss2 with β of 2.5 also reaches a D-Return of 1.322 over 5 years, significantly above
the -0.024 of the MSE loss. The custom loss function significantly over-performs MSE in
terms of expected return.
With D-VaR above 1 across the table, all loss functions achieve a risk reduction compared
to the buy & hold strategy. The most efficient loss function for the risk reduction is AdjLoss2
with β of 2.25.

Table 3. Average D ratio, D-Return and D-VaR per loss function (the higher the better)

5.3. Distribution of the D ratios and Area Under Curve (AUC)

The custom loss functions achieve an average performance significantly above the MSE
loss. To reinforce these positive results, we test whether the average result is also consistent
for most individual assets. Plotting the cumulative distribution of D ratios (Graph. 5) is a way
to assess the consistency of the results across the 105 stocks. The graph demonstrates visually
that MSE is the worst loss function with more D ratios below 1 and even below 0.
This plot of the cumulative distribution can be translated into a metric of Area Under
Curve, presented in Table 4. The AUC of D ratio can also be computed for D-return and D-
VaR.
Table 4. Normalized Area Under Curve of D ratio per loss function (the lower the better)

15

Electronic copy available at: https://ssrn.com/abstract=3973086

 The same custom loss function AdjLoss2 with β of 2.5 provides the lowest AUC for the D
ratio, demonstrating that the custom loss function is also more effective than MSE for most of
the individual stocks.
The AUC of D-Return of AdjLoss2 with β of 2.5 confirms that the custom loss function is
more effective than MSE to produce effective returns for most stocks. As AUC is below 1,
the loss function also overperforms the buy & hold strategy.
Graph 5. Cumulative distribution of D ratio per loss function

6. Conclusion and perspectives

In this final section, we conclude on the results from our analysis. We then draw some
perspective for future research.

6.1. Conclusion
MSE is the most common loss function applied in ML and DL. While it is easy to apply,
MSE delivers sub-optimal results once compared with asymmetric custom loss functions for
algorithms predicting asset returns, as the consequence of the error in prediction does not
deliver symmetrical consequences in terms of effective realised return.
Customization of the loss function to render the asymmetric consequence of the error in
prediction is an easy way to significantly improve the results of simple algorithms aiming at
predicting results. Not only do most custom loss functions perform much better than algorithms
with MSE, but they also manage to achieve better results than a buy & hold strategy, a result
that MSE never achieved with our MLP algorithm. Depending on the risk aversion of the
investor and on the benchmark strategy for reference, various customizations can be
contemplated. Lloss function AdjLoss2 appears to be among the best performers, not only in
terms of risk-adjusted return metric (D ratio) but also in terms of equilibrium between the
effective return (D-return) and the effective risk-reduction (D-VaR). Eventually, this loss

16

Electronic copy available at: https://ssrn.com/abstract=3973086

function provides a relative stability of the performance over time, when measured over several
sub-periods.

6.2. Perspectives.
Loss function is a key component of any ML algorithm and a key input for computing the
gradient descent. We demonstrate the advantage of tailoring the loss function with a simple
deep learning model. Three possible next steps could be implemented, that would generalise
our results: (i) testing the algorithm with other types of assets (bonds, ETFs, commodities,
crypto-currencies, …), (ii) testing the superiority of custom loss functions with more complex
algorithms (LSTM, CNN and ResNet) for performing the same task of predicting asset returns.
Eventually, (iii) generalising the principle of custom loss function could possibly be efficiently
applied, mutatis mutandis, to some Reinforcement Learning (RL) algorithms (like an on-policy
actor-critic PPO model).

Declaration of Competing Interest

The author declares that he has no known competing financial interests or personal
relationships that could have appeared to influence the work reported in this paper.

Funding
This research did not receive any specific grant from funding agencies in the public,
commercial, or not-for-profit sectors.

Warning
This research is for scientific purposes only and is not intended to support any model of
investment or trading.

References
Abe, M., Nakayama, H., 2018. Deep learning for forecasting stock returns in the cross-
section. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect.
Notes Bioinformatics) 10937 LNAI, 273–284. https://doi.org/10.1007/978-3-319-93034-
3_22
Abroyan, N., 2017. Neural Networks for Financial Market Risk Classification. Front. Signal
Process. 1, 62–66. https://doi.org/10.22606/fsp.2017.12002
Ahmed, S., Hassan, S.-U., Aljohani, N.R., Nawaz, R., 2020. FLF-LSTM: A novel prediction

17

Electronic copy available at: https://ssrn.com/abstract=3973086

 system using Forex Loss Function. Appl. Soft Comput. J. 97, 106780.
https://doi.org/10.1016/j.asoc.2020.106780
Barton, S., Alakkari, S., O’Dwyer, K., Ward, T., Hennelly, B., 2021. Convolution Network
with Custom Loss Function for the Denoising of Low SNR Raman Spectra. Sensors
2021, Vol. 21, Page 4623 21, 4623. https://doi.org/10.3390/S21144623
Bhandari, B., Alsadoon, A., Prasad, P.W.C., Abdullah, S., Haddad, S., 2020. Deep learning
neural network for texture feature extraction in oral cancer: enhanced loss function.
Multimed. Tools Appl. 2020 7937 79, 27867–27890. https://doi.org/10.1007/S11042-
020-09384-6
Börjesson, L., Singull, M., 2020. Forecasting financial time series through causal and dilated
convolutional neural networks. Entropy 22, 1–20. https://doi.org/10.3390/e22101094
Borovkova, S., Tsiamas, I., 2019. An ensemble of LSTM neural networks for high-frequency
stock market classification. J. Forecast. 38, 600–619. https://doi.org/10.1002/for.2585
Bustos, O., Pomares-Quimbaya, A., 2020. Stock market movement forecast: A Systematic
review. Expert Syst. Appl. 156. https://doi.org/10.1016/j.eswa.2020.113464
Chen, J., Wu, W., Tindall, M.L., 2016. Hedge Fund Return Prediction and Fund Selection: A
Machine-Learning Approach. Financ. Ind. Stud. Dep. Dallas Fed.
Chen, L., Pelger, M., Zhu, J., 2020. Deep Learning in Asset Pricing. SSRN Electron. J.
https://doi.org/10.2139/ssrn.3350138
Chen, X., Yu, R., Ullah, S., Wu, D., Li, Z., Li, Q., Qi, H., Liu, J., Liu, M., Zhang, Y., 2021. A
novel loss function of deep learning in wind speed forecasting. Energy.
https://doi.org/10.1016/j.energy.2021.121808
Christoffersen, P., Diebold, F.X., 1996. Further Results on Forecasting and Model Selection
Under Asymmetric Loss Author ( s ): Peter F . Christoffersen and Francis X . Diebold
Source : Journal of Applied Econometrics , Vol . 11 , No . 5 , Special Issue :
Econometric Published by : Wiley Stable. J. Appl. Econom. 11, 561–571.
Cuturi, M., Blondel, M., 2017. Soft-DTW: a Differentiable Loss Function for Time-Series, in:
34 Th International Conference on Machine Learning. PMLR, pp. 894–903.
Dessain, J., 2022. Machine learning models predicting returns: Why most popular
performance metrics are misleading and proposal for an efficient metric. Expert Syst.
Appl. 199, 116970. https://doi.org/10.1016/j.eswa.2022.116970
Ding, X., Zhang, Y., Liu, T., Duan, J., 2015. Deep Learning for Event-Driven Stock
Prediction Xiao. J. Sci. Ind. Res. (India). 2327–2333.
Dingli, A., Fournier, K.S., 2017. Financial time series forecasting - a deep learning approach.
Int. J. Mach. Learn. Comput. 7, 118–122. https://doi.org/10.18178/ijmlc.2017.7.5.632
Dress, K., Lessmann, S., Von Mettenheim, H.-J., 2018. Residual value forecasting using
asymmetric cost functions. Int. J. Forecast. 34, 551–565.
https://doi.org/10.1016/j.ijforecast.2018.01.008
Ebert-Uphoff, I., Lagerquist, R., Hilburn, K., Lee, Y., Haynes, K., Stock, J.C., Kumler, C.,
Stewart, J.Q., 2021. CIRA Guide to custom loss functions for neural networks in
environmental sciences - Version 1. arXiv 2106.09757.
Elliot, A., Hsu, C.H., 2017. Time Series Prediction : Predicting Stock Price. arXiv
1710.05751.

18

Electronic copy available at: https://ssrn.com/abstract=3973086

Eramo, V., Giacinto Lavacca, F., Catena, T., Jaime, P., Salazar, P., 2021. Application of a
Long Short Term Memory neural predictor with asymmetric loss function for the
resource allocation in NFV network architectures. Comput. Networks.
https://doi.org/10.1016/j.comnet.2021.108104
Feng, G., He, J., Polson, N.G., 2018a. Deep learning for predicting asset returns. arXiv
1804.09314 1–23.
Feng, G., Polson, N.G., Xu, J., 2018b. Deep Learning in Characteristics-Sorted Factor
Models. arXiv 1805.01104 1–41.
Fischer, T., Krauss, C., 2018. Deep learning with LSTM networks for Financial Market
Predictions. Eur. J. Oper. Res. 270, 1–34.
Gu, S., Kelly, B., Xiu, D., 2020. Empirical Asset Pricing via Machine Learning. Rev. Financ.
Stud. 33, 2223–2273. https://doi.org/10.1093/rfs/hhaa009
Hansun, S., Young, J.C., 2021. Predicting LQ45 financial sector indices using RNN-LSTM.
J. Big Data 8. https://doi.org/10.1186/s40537-021-00495-x
Hao, Y., Gao, Q., 2020. Predicting the trend of stock market index using the hybrid neural
network based on multiple time scale feature learning. Appl. Sci. 10.
https://doi.org/10.3390/app10113961
Henrique, B.M., Sobreiro, V.A., Kimura, H., 2018. Stock price prediction using support
vector regression on daily and up to the minute prices. J. Financ. Data Sci. 4, 183–201.
https://doi.org/10.1016/j.jfds.2018.04.003
Huang, J., Chai, J., Cho, S., 2020. Deep learning in finance and banking: A literature review
and classification. Front. Bus. Res. China 14. https://doi.org/10.1186/s11782-020-00082-
6
Hwang, S., Knight, J., Stephen, S., 2001. Forecasting Nonlinear Functions of Returns Using
LINEX Loss Functions. Ann. Econ. Financ. 2, 187–213.
Kao, L.J., Chiu, C.C., Lu, C.J., Yang, J.L., 2013. Integration of nonlinear independent
component analysis and support vector regression for stock price forecasting.
Neurocomputing 99, 534–542. https://doi.org/10.1016/j.neucom.2012.06.037
Lim, B., Zohren, S., Roberts, S., 2019. Enhancing Time-Series Momentum Strategies Using
Deep Neural Networks. J. Financ. Data Sci. 1, 19–38.
https://doi.org/10.3905/jfds.2019.1.015
Liu, H., 2018. Leveraging Financial News for Stock Trend Prediction with Attention-Based
Recurrent Neural Network. arXiv1811.06173.
Long, W., Lu, Z., Cui, L., 2019. Deep learning-based feature engineering for stock price
movement prediction. Knowledge-Based Syst. 164, 163–173.
https://doi.org/10.1016/j.knosys.2018.10.034
Ma, Y., Han, R., Wang, W., 2021. Portfolio optimization with return prediction using deep
learning and machine learning. Expert Syst. Appl. 165.
https://doi.org/10.1016/j.eswa.2020.113973
Mehtab, S., Sen, J., 2020. A Time Series Analysis-Based Stock Price Prediction Using
Machine Learning and Deep Learning Models. arXiv 2004.11697 1–46.
https://doi.org/10.13140/RG.2.2.14022.22085/2
Meng, T.L., Khushi, M., 2019. Reinforcement learning in financial markets. Data 4, 1–17.

19

Electronic copy available at: https://ssrn.com/abstract=3973086

 https://doi.org/10.3390/data4030110
Nikou, M., Mansourfar, G., Bagherzadeh, J., 2019. Stock price prediction using DEEP
learning algorithm and its comparison with machine learning algorithms. Intell. Syst.
Accounting, Financ. Manag. 26, 164–174. https://doi.org/10.1002/isaf.1459
Ozbayoglu, A.M., Gudelek, M.U., Sezer, O.B., 2020. Deep learning for financial
applications : A survey. arXiv 2002.05786.
Parida, A.K., Bisoi, R., Dash, P.K., 2016. Chebyshev polynomial functions based locally
recurrent neuro-fuzzy information system for prediction of financial and energy market
data. J. Financ. Data Sci. 2, 202–223. https://doi.org/10.1016/J.JFDS.2016.10.001
Patton, A.J., Timmermann, A., 2007. Properties of optimal forecasts under asymmetric loss
and nonlinearity. J. Econom. 140, 884–918.
https://doi.org/10.1016/j.jeconom.2006.07.018
Persio, L. Di, Honchar, O., 2017. Recurrent neural networks approach to the financial
forecast of Google assets. Int. J. Math. Comput. Simul. 11, 7–13.
Qin, Y., Song, D., Cheng, H., Cheng, W., Jiang, G., Cottrell, G.W., 2017. A dual-stage
attention-based recurrent neural network for time series prediction. IJCAI Int. Jt. Conf.
Artif. Intell. 0, 2627–2633. https://doi.org/10.24963/ijcai.2017/366
Raein Hashemi, S., Sadegh Mohseni Salehi, S., Member, Student, Erdogmus, D., Member,
Senior, Prabhu, S.P., Warfield, S.K., Gholipour, A., 2019. Asymmetric Loss Functions
and Deep Densely-Connected Networks for Highly-Imbalanced Medical Image
Segmentation: Application to Multiple Sclerosis Lesion Detection 7, 2169–3536.
https://doi.org/10.1109/ACCESS.2018.2886371
Rout, A.K., Dash, P.K., Dash, R., Bisoi, R., 2017. Forecasting financial time series using a
low complexity recurrent neural network and evolutionary learning approach. J. King
Saud Univ. - Comput. Inf. Sci. 29, 536–552.
https://doi.org/10.1016/J.JKSUCI.2015.06.002
Roy Choudhury, A., Abrishami, S., Turek, M., Kumar, P., 2020. Enhancing profit from stock
transactions using neural networks. AI Commun. 33, 75–92.
https://doi.org/10.3233/AIC-200629
Sezer, O.B., Gudelek, M.U., Ozbayoglu, A.M., 2020. Financial time series forecasting with
deep learning: A systematic literature review: 2005–2019. Appl. Soft Comput. J. 90.
https://doi.org/10.1016/j.asoc.2020.106181
Shen, G., Tan, Q., Zhang, H., Zeng, P., Xu, J., 2018. Deep Learning with Gated Recurrent
Unit Networks for Financial Sequence Predictions. Procedia Comput. Sci. 131, 895–903.
https://doi.org/10.1016/J.PROCS.2018.04.298
Sim, H.S., Kim, H.I., Ahn, J.J., 2019. Is Deep Learning for Image Recognition Applicable to
Stock Market Prediction? Complexity 2019. https://doi.org/10.1155/2019/4324878
Singh, B.K., 2014. Loss Functions in Financial Sector: An Overview. Asian J. Math. Stat. 8,
35–45. https://doi.org/10.3923/ajms.2015.35.45
Song, D., Busogi, M., Chung Baek, A.M., Kim, N., 2020. Forecasting stock market index
based on pattern-driven long short-term memory. Econ. Comput. Econ. Cybern. Stud.
Res. 54, 25–41. https://doi.org/10.24818/18423264/54.3.20.02
Tavakoli, M., Doosti, H., 2021. Forecasting the Tehran Stock market by Machine Learning
Methods using a New Loss Function. Adv. Math. Financ. Appl. 6, 194–205.

20

Electronic copy available at: https://ssrn.com/abstract=3973086

 https://doi.org/10.22034/AMFA.2020.1896273.1399
Thakkar, A., Chaudhari, K., 2021. A comprehensive survey on deep neural networks for
stock market: The need, challenges, and future directions. Expert Syst. Appl. 177,
114800. https://doi.org/10.1016/j.eswa.2021.114800
Thakkar, A., Chaudhari, K., 2020. CREST: Cross-Reference to Exchange-based Stock Trend
Prediction using Long Short-Term Memory. Procedia Comput. Sci. 167, 616–625.
https://doi.org/10.1016/J.PROCS.2020.03.328
Tsantekidis, A., Passalis, N., Tefas, A., Kanniainen, J., Gabbouj, M., Iosifidis, A., 2020.
Using Deep Learning for price prediction by exploiting stationary limit order book
features. Appl. Soft Comput. J. 93. https://doi.org/10.1016/j.asoc.2020.106401
Wen, M., Li, P., Zhang, L., Chen, Y., 2019. Stock market trend prediction using high-order
information of time series. IEEE Access 7, 28299–28308.
https://doi.org/10.1109/ACCESS.2019.2901842
Wen, Q., Yang, Z., Song, Y., Jia, P., 2010. Automatic stock decision support system based on
box theory and SVM algorithm. Expert Syst. Appl. 37, 1015–1022.
https://doi.org/10.1016/j.eswa.2009.05.093
Wu, J., Wang, Y.-G., Tian, Y.-C., Burrage, K., Cao, T., 2021. Support vector regression with
asymmetric loss for optimal electric load forecasting. Energy.
https://doi.org/10.1016/j.energy.2021.119969
Yan, X., Weihan, W., Chang, M., 2021. Research on financial assets transaction prediction
model based on LSTM neural network. Neural Comput. Appl. 33, 257–270.
https://doi.org/10.1007/s00521-020-04992-7
Yun, H., Lee, M., Kang, Y.S., Seok, J., 2020. Portfolio management via two-stage deep
learning with a joint cost. Expert Syst. Appl. 143, 113041.
https://doi.org/10.1016/j.eswa.2019.113041
Zellner, A., 1986. Bayesian estimation and prediction using asymmetric loss functions. J. Am.
Stat. Assoc. 81. https://doi.org/10.1080/01621459.1986.10478289
Zhou, X., Pan, Z., Hu, G., Tang, S., Zhao, C., 2018. Stock Market Prediction on High-
Frequency Data Using Generative Adversarial Nets. Math. Probl. Eng. 2018.
https://doi.org/10.1155/2018/4907423

21

Electronic copy available at: https://ssrn.com/abstract=3973086

Appendices
Appendix 1. Loss functions applied by author
Author Loss functions
Abe and Nakayama (2018) Cross-entropy (classification)
Abroyan (2017) Binary cross-entropy (classification)
Börjesson and Singull (2020) MAPE (regression)
Borovkova & Tsiamas (2019) Cross-entropy (classification)
Chen et al. (2016) MSE (regression)
Chen et al. (2020) Weighted mean pricing error (GAN)
Ding et al. (2015) Margin loss (classification)
Dingli and Fournier (2017) Margin loss (classification)
Elliot and Hsu (2017) Unspecified
Feng, He, et al. (2018) MSE (regression)
Feng, Polson, et al. (2018) Economic-driven Fama-French
Fischer & Krauss (2018) Cross-entropy (classification)
Gu et al. (2020) MSE, MAE and Huber11 (regression)
Hansun & Young (2021) MSE (regression)
Hao & Gao (2020) Binary cross-entropy (classification)
Henrique et al. (2018) Vapnik (SVM)
Kao et al. (2013) Cite Vapnik (SVM)
Lim et al. (2019) MSE + sharpe (regression)
Liu (2018) Cross-entropy (classification)
Long et al. (2019) Cross-entropy (classification)
Ma et al. (2021) MAE (regression)
Mehtab and Sen (2020) MAE (regression)
Nikou et al. (2019) MSE (regression), cross-entropy, Hinge (classification)
Parida et al. (2016) MSE (regression)
Persio and Honchar (2017) Log proba (classification)
Qin et al. (2017) RMSE (regression)
Rout et al. (2017) MSE (regression)
Roy Choudhury et al. (2020) MSE (regression)
Shen et al. (2018) Cross-entropy versus margin-based loss for SVM
(classification)
Sim et al. (2019) MSE (regression)
Song et al. (2020) MSE (regression)
Thakkar and Chaudhari (2020) MSE (regression)
Tsantekidis et al. (2020) Categorical cross-entropy (classification)
Wen et al. (2019) MSE (regression)
Wen et al. (2010) Huber (for SVM)
Yan et al. (2021) Categorical cross-entropy (classification)
Yun et al. (2020) Joint cost function

11
Huber loss function as a mix of MSE and MAE

22

Electronic copy available at: https://ssrn.com/abstract=3973086

 Zhou et al. (2018) Combined MSE + directional loss (regression)
Appendix 2. List of stocks
# RIC Asset # RIC Asset
1 AAPL Apple Inc 44 AGN.AS Aegon NV
2 AIG AIG 45 AKZA.AS Akzo Nobel NV
3 AMGN Amgen 46 ASML.AS ASML Holding NV
4 AXP American Express 47 DSM.AS Koninklijke DSM NV
5 BA Boeing 48 HEIO.AS Heineken NV
6 BAC Bank of America 49 INGA.AS ING Groep NV
7 C Citigroup 50 KPN.AS Koninklijke KPN NV
8 CAT Caterpillar 51 MT.AS ArcelorMittal SA
9 CRM Salesforces.com 52 PHIA.AS Koninklijke Philips NV
10 CSCO Cisco Systems 53 RAND.AS Randstad NV
11 CVX Chevron 54 RDSa.AS Royal Dutch Shell PLC
12 DD Dupont de Nemours 55 UNA.AS Umicore SA
13 DIS Walt Disney 56 URW.AS Unibail-Rodamco-Westfield
14 FL Foot Locker 57 WKL.AS Wolters Kluwer NV
15 GE GE 58 ABI.BR Anheuser Busch Inbev NV
16 GS Goldman Sachs 59 ACKB.BR Ackermans & Van Haaren NV
17 GT Goodyear Tire 60 AGS.BR Ageas SA
18 HD Home Depot 61 BAR.BR Barco NV
19 HON Honeywell International 62 COFB.BR Cofinimmo SA
20 HPQ Hewlett Packard 63 COLR.BR Colruyt NV
21 IBM IBM 64 GBLB.BR Groep Brussel Lambert NV
22 INTC Intel Corp 65 KBC.BR KBC Groep NV
23 IP International Paper 66 PROX.BR Proximus NV
24 JNJ Johnson & Johnson 67 SOF.BR Sofina SA
25 JPM JP Morgan 68 SOLB.BR Solvay SA
26 KO Coca Cola 69 TNET.BR Telenet Group Holding NV
27 MCD McDonald's 70 UCB.BR Ucb SA
28 MDLZ Mondelez Intl 71 UMI.BR Umicore SA
29 MMM 3M 72 WDP.BR Warehouses de Pauw
30 MO Altria Group 73 AC.PA Accor SA
31 MRK Merck & Co 74 ACA.PA Credit Agricole SA
32 MSFT Microsoft 75 AI.PA L'Air Liquide SA
33 NKE Nike 76 AIR.PA Airbus SE
34 PFE Pfizer 77 ATO.PA Atos SE
35 PG Procter & Gamble 78 BN.PA Danone SA
36 RTX Raytheon Tech 79 BNP.PA BNP Paribas SA
37 T AT&T 80 CA.PA Carrefour SA
38 TRV Travelers Companies 81 CAP.PA Capgemini SE
39 UNH UnitedHealth Group 82 DG.PA Vinci SA
40 VZ Verizon Communications 83 DSY.PA Dassault Systemes SE
41 WBA Walgreens Boots Alliance 84 EN.PA Bouygues SA
42 WMT Walmart 85 ENGI.PA Engie SA
43 XOM Exxon Mobil 86 FP.PA Total SA

23

Electronic copy available at: https://ssrn.com/abstract=3973086

 # RIC Asset
87 HO.PA Thales SA
88 KER.PA Kering SA
89 MC.PA LVMH
90 ML.PA Michelin SA
91 OR.PA L'Oreal SA
92 ORA.PA Orange SA
93 PUB.PA Publicis Groupe SA
94 RI.PA Pernod Ricard SA
95 RMS.PA Hermes International SCA
96 RNO.PA Renault SA
97 SAF.PA Safran SA
98 SAN.PA Sanofi SA
99 SGO.PA Compagnie de Saint Gobain SA
100 STM.PA STMicroelectronics NV
101 SU.PA Schneider Electric SE
102 TEP.PA Teleperformance SE
103 UG.PA Peugeot SA
104 VIE.PA Veolia Environnement SA
105 VIV.PA Vivendi SA

Appendix 3. Hardware and software

Hardware :
In order to assess the reproducibility of the models, two different hardware were used to
perform the computation.

HARDWARE PC1 PC2

Processor AMD Ryzen 9 3950 X Intel i7-8700
CPU DDR4 DDR3
GPU Nvidia RTX 3090 24GB Nvidia GTX 1060i 6GB
RAM 64 GB 16 GB

Software :
The two computers run similar software with, sometimes, marginal differences in releases.
SOFTWARE PC1 PC2
OS Windows 10 Pro Windows 10 Pro
Anaconda 1.9.12 1.9.12
Spyder 4.1.5 4.1.5
Python 3.8 3.7

24

Electronic copy available at: https://ssrn.com/abstract=3973086

 Cuda 11.0.221 10.2.89
CUDNN 7.6.5

Pytorch 1.7.0 1.7.0

Tensorflow 2.1.0 2.1.0
Tensorflow-GPU 2.3.0 2.3.0
scikit-learn 0.23.2 0.23.2
Stable Baselines3 1.0 1.0

Numpy 1.19.2 1.19.2

Pandas 1.1.3 1.1.3
Matplotlib 3.3.2 3.3.2

Sqlite 3.33.0 3.33.0

pyts 0.11.0 0.11.0

Neptune Client 0.9.5 0.9.4

Contrib 0.27.1 0.27.1

In order to secure reproducibility to the larger extent possible, we applied several strategies :
- Seed defined for Python, Numpy, TensorFlow and/or Pytorch (both CPU and GPU)
- Deterministic backend forced for CUDNN
- debug environment variable CUBLAS_WORKSPACE_CONFIG defined to
":4096:8"

25

Electronic copy available at: https://ssrn.com/abstract=3973086