Simon, Frederik; Weibels, Sebastian; Zimmermann, Tom

Working Paper
Deep parametric portfolio policies

CFR Working Paper, No. 23-01

Provided in Cooperation with:

Centre for Financial Research (CFR), University of Cologne

Suggested Citation: Simon, Frederik; Weibels, Sebastian; Zimmermann, Tom (2023) : Deep parametric
portfolio policies, CFR Working Paper, No. 23-01, University of Cologne, Centre for Financial
Research (CFR), Cologne

This Version is available at:

https://hdl.handle.net/10419/270745

Standard-Nutzungsbedingungen: Terms of use:

Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Documents in EconStor may be saved and copied for your personal
Zwecken und zum Privatgebrauch gespeichert und kopiert werden. and scholarly purposes.

Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle You are not to copy documents for public or commercial purposes, to
Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich exhibit the documents publicly, to make them publicly available on the
machen, vertreiben oder anderweitig nutzen. internet, or to distribute or otherwise use the documents in public.

Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen If the documents have been made available under an Open Content
(insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, Licence (especially Creative Commons Licences), you may exercise
gelten abweichend von diesen Nutzungsbedingungen die in der dort further usage rights as specified in the indicated licence.
genannten Lizenz gewährten Nutzungsrechte.
 CFR Working Paper NO. 23-01

Deep Parametric Portfolio Policies

F. Simon • S. Weibels
• T. Zimmermann
 Deep Parametric Portfolio Policies*

Frederik Simon† Sebastian Weibels‡ Tom Zimmermann§

First Version: July 2022; This Version: February 2023

Abstract
We directly optimize portfolio weights as a function of firm characteristics via deep neural
networks by generalizing the parametric portfolio policy framework. Our results show that
network-based portfolio policies result in an increase of investor utility of between 30 and 100
percent over a comparable linear portfolio policy, depending on whether portfolio restrictions
on individual stock weights, short-selling or transaction costs are imposed, and depending
on an investor’s utility function. We provide extensive model interpretation and show that
network-based policies better capture the non-linear relationship between investor utility and
firm characteristics. Improvements can be traced to both variable interactions and non-linearity
in functional form. Both the linear and the network-based approach agree on the same
dominant predictors, namely past return-based firm characteristics.

JEL classification: G11, G12, C58, C45

Keywords: Portfolio Choice, Machine Learning, Expected Utility

* We thank Bryan Kelly, Victor DeMiguel, Christian Fieberg, Alexander Klos, Simon Rottke, Fabricius Somogyi
(discussant), Bastidon Cécile (discussant) and participants at the Research in Behavioral Finance Conference (RBFC), the
Cardiff Fintech Conference, the 2022 New Zealand Finance Meeting (NZFM) as well as the Paris Financial Management
Conference (PFMC) for helpful comments and suggestions.
† University of Cologne, Department of Business Administration and Corporate Finance, simon.frederik@wiso.uni-

koeln.de
‡ University of Cologne, Department of Business Administration and Bank Management, weibels@wiso.uni-koeln.de
§ University of Cologne, Institute for Econometrics and Statistics and Center for Financial Research,

tom.zimmermann@uni-koeln.de

1
1 Introduction

Consider the formidable problem of an investor who wants to choose an optimal asset allocation

within her equity portfolio. The literature provides her with a few options: She can opt for a

traditional Markowitz approach (Markowitz, 1952) that requires estimation of expected returns,

variances and covariances with the number of moments to estimate escalating quickly. At the

other end of the spectrum, she might estimate a low-dimensional parametric portfolio policy

(PPP) (Brandt et al., 2009) but the linear model might not provide sufficient flexibility. She can

also consult a large literature that relates characteristics to expected returns but even studies that

consider a multitude of firm-level characteristics (e.g. Gu et al., 2020) only investigate expected

returns and do not speak to risk as perceived by different investors’ objective functions.

We provide a general solution to the portfolio optimization challenge. In short, we combine

the parametric portfolio policy approach that is well-suited to estimate portfolio weights for any

utility function with the flexibility of feed-forward networks from the machine learning literature.

The resulting approach that we label Deep Parametric Portfolio Policy (DPPP) is well-suited to

accommodate flexible non-linear and interactive relationships between portfolio weights and stock

characteristics, to integrate different utility functions, to deal with leverage or portfolio weight

constraints, and to incorporate transaction costs.

Our results are fourfold. First, our model improves significantly over a standard linear

parametric portfolio policy. Utility gains range from around 30% to 100% depending on model

specification and the incorporation of constraints. Such gains are not restricted to only particular

time periods and can be attributed to the fact that the relationship between firm characteristics

and investor utility is non-linear. Second, in our benchmark model, past return-based stock

characteristics turn out to be more relevant to the portfolio policy than accounting-based charac-

teristics. However, in line with extant literature (DeMiguel et al., 2020; Jensen et al., 2022), the

relevance of return-based characteristics decreases when we model transaction costs explicitly in

the objective function. Third, utility gains arise for a variety of investors’ utility functions that we

consider. While our benchmark investor is a classical mean-variance optimizer, our setup easily

accommodates other utility functions. We also investigate deep parametric portfolio policies for

the case of constant relative risk aversion and for loss aversion, and we find substantial utility

2
gains in all cases. Fourth, we show that both non-linearity in variables (i.e. variable interactions)

and non-linearity in functional form account for the differences between the estimated weights of

the linear parametric portfolio policy and our model.

In essence, our model can be interpreted as a generalization of the linear parametric portfolio

policy approach. More specifically, we allow portfolio weights to be of one of the arguably most

flexible forms - a neural network. This represents a significant conceptual deviation from linear

parametric portfolio policies in two ways: First, by replacing the linear specification by a neural

network, we allow the relation between firm characteristics and weights to be non-linear and we

allow for potential interactions of firm characteristics. The literature on using machine learning

methods to predict future returns shows that such flexibility is relevant to model the relation

between firm characteristics and future returns, and can lead to substantial improvements over

less flexible specifications (Moritz and Zimmermann, 2016; Freyberger et al., 2020; Gu et al., 2020).

It is conceivable that such flexibility will also help to model the relation between portfolio weights

and firm characteristics. Second, this flexibility comes at the cost of having to estimate a model

with a high-dimensional parameter vector. As such, it deviates from the original motivation

of the parametric portfolio policy literature that aimed to reduce portfolio optimization to a

low-dimensional problem in which only a small number of coefficients need to be estimated. Our

benchmark model has around 5,700 parameters compared to the three parameters that need to

be estimated in the application of Brandt et al. (2009). Nevertheless, Kelly et al. (2022) argue

that model complexity is a virtue for return prediction, and our approach can be viewed as an

exploration of that point in the context of parametric portfolio policies.

Building on Brandt et al. (2009), we begin with a benchmark case of a largely unrestricted

portfolio policy. In the benchmark case, an investor who optimizes mean-variance utility can

take long and short positions with the only restriction that absolute individual stock positions

cannot exceed three percent of the overall portfolio. Other aspects of the optimization remain

unrestricted, in particular, the investor does not take into account transaction costs or short-selling

constraints.

In the benchmark case, a network-based portfolio policy can increase investor utility by about

100% relative to a linear portfolio policy but also incurs higher turnover. Both portfolio policies

take comparably large positions in individual stocks but the network-based policy has turnover

3
that is almost twice as large. We find that the difference in turnover can be traced to the network-

based policy putting larger weight on past-return based characteristics that imply higher turnover,

such as e.g. short-term reversal.

We then investigate network-based portfolio policies in more realistic contexts that restrict

investors in various ways. In particular, we explore results for the case in which an investor cannot

take any short positions and for the case in which transaction costs and leverage are part of the

optimization problem. In both cases, we find that network-based policies yield higher utility than

a linear portfolio policy, with increases between 30% and 40%. For constrained portfolio policies

the importance of past-return based characteristics decreases while still being among the most

important characteristics. This matches the results of DeMiguel et al. (2020) who find that more

characteristics matter under transaction costs.

Moving beyond our benchmark mean-variance investor, we explore different investor prefer-

ences: First, we show that utility gains relative to a benchmark portfolio occur for mean-variance

utility optimizers with different degrees of risk aversion. We find larger utility gains for less

risk averse investors and lower gains for more risk averse investors, consistent with our finding

that estimated portfolio policies for more risk averse investors take less extreme positions and

hold more diversified portfolios. Second, we also find that utility gains are not restricted to

mean-variance utility investors. We find similar results when we consider constant relative risk

aversion or loss aversion.

Overall, our contribution can be summarized as providing a general solution to the parametric

portfolio policy problem that combines recent advances in combining structural economic prob-

lems and machine learning methods (Farrell et al., 2021; Kelly et al., 2022). Our setup seamlessly

incorporates non-linearities in parameters and across firm characteristics. We also demonstrate

how constraints on leverage or portfolio weights can be easily added via customization of the

statistical loss function. Lastly, realistic estimates of transaction costs can be taken into account as

an additional constraint on the optimization problem.

4
1.1 Related Literature

Our work relates to three different strands of the literature. First, we add to a growing literature

that explores the potential of machine learning algorithms in finance (e.g. Heaton et al., 2017; Gu

et al., 2020; Bianchi et al., 2020; Kelly et al., 2022). Studies in this literature typically consider a

prediction task (e.g. predicting stock returns), and optimize a standard statistical loss function

such as the mean squared error (or a related distance metric) between the actual and predicted

values. Predicted values are used to construct portfolio weights (e.g. Gu et al., 2020). In contrast,

we optimize a utility function instead of a common loss function and model portfolio weights

directly as a function of firm characteristics. The use of machine learning algorithms to estimate

coefficients of structural models (in our case portfolio weights) as flexible functions has also been

proposed recently by Farrell et al. (2021).

Second, we extend the literature on one-step portfolio optimization. Specifically, we extend the

parametric portfolio approach by Brandt et al. (2009). While Brandt et al. (2009) argue that it may

be worthwhile to consider non-linear functions and interactions in weight modeling, subsequent

papers that have implemented and extended parametric portfolio policies parameterize portfolio

weights as a linear function of firm characteristics (e.g. Hjalmarsson and Manchev, 2012; Ammann

et al., 2016). DeMiguel et al. (2020) incorporate transaction costs, a larger set of firm characteristics,

and statistical regularization but also stay within the linear framework. Our deep parametric

portfolio policy replaces the linear model with a feed-forward neural network that accounts for

both non-linearity and possible interactions of firm characteristics. In addition, we use a larger

set of firm characteristics than previous studies and explore different regularization techniques

for both the linear and deep parametric portfolio policies. Alternative, (machine learning-based)

one-step portfolio optimization approaches include Cong et al. (2021), Butler and Kwon (2021),

Uysal et al. (2021), Chevalier et al. (2022) and Jensen et al. (2022). Each of these differs from ours in

one or more aspects. First and foremost, in contrast to any of these we generalize the approach of

Brandt et al. (2009) and explicitly analyze differences between a linear and non-linear specification.

In addition, Cong et al. (2021) use a general reinforcement learning - based approach and sort

stocks into portfolios to maximize Sharpe ratios while our feed-forward network directly optimizes

continuous portfolio weights for various investor utility functions. Butler and Kwon (2021) show

5
that it is possible to integrate regression-based return predictions into the portfolio optimization

by means of a two-layer neural network, one layer resembling the return prediction and one layer

resembling the weight optimization. However, their results are restricted to a mean-variance

setting, while our approach is flexibly applicable to any type of investor preference. Moreover, our

empirical analysis is about modeling portfolios of stocks based on stock characteristics, whereas

they empirically assess their models on simulated data and commodity future markets. Chevalier

et al. (2022) derive optimal in-sample weights based on investor preferences and subsequently

predict these weights conditional on covariates. This is conceptually different from our approach,

primarily because we do not require the preprocessing step of computing the optimal in-sample

weights. Jensen et al. (2022) take a different angle. They aim to specifically tackle the issue of

integrating transaction costs into mean-variance portfolio optimization with machine learning.

They outline different approaches to do so, inter alia, a ML-based one-step approach. However,

rather than extending the approach by Brandt et al. (2009) as we do, they derive a closed form

solution to the issue and implement it empirically by using random feature regressions. Moreover,

their focus in terms of interpreting the empirical relations lays on comparing different approaches

to achieve the aforementioned underlying aim of integrating transaction costs, as well as the

importance of features in this setting. We, in contrast, also shed light onto how non-linearities

contribute to portfolio optimization.

Finally, we relate to the literature that examines which firm characteristics are jointly significant

in explaining expected returns (Fama and French, 2008; Green et al., 2017; Freyberger et al., 2020).

While all of these studies focus on cross-sectional regression models with extensions, Gu et al.

(2020) find that neural networks perform best in predicting mean returns for a large number of

firm characteristics. Our portfolio approach using neural networks considers all moments of the

return distribution beyond the expected return if they are relevant to an investor’s utility function.

Most of this literature ignores various real world constraints such as transaction costs (with

Novy-Marx and Velikov (2016), DeMiguel et al. (2020) and Jensen et al. (2022) being important

exceptions) or weight constraints, whereas we show how our model allows us to seamlessly

integrate transaction costs or other constraints.

6
2 Model

2.1 Expected Utility Framework and Parametric Portfolio Policies

The starting point of our framework is the parametric portfolio policy model in Brandt et al.

(2009). Consider a universe of Nt stocks that an investor can invest in at each month t ∈ T. Each

stock i is associated with a vector of firm characteristics xi,t and a return ri,t+1 from date t to t + 1.

An investor’s objective is to maximize the conditional expected utility of future portfolio returns

r p,t+1 : " !#
Nt
∑ wi,t ri,t+1
 
max Et u(r p,t+1 ) = Et u , (1)
N
{wi,t }i=t1 i =1

where wi,t is the weight of stock i in the portfolio at date t and u(·) denotes the respective utility

function.

Instead of directly deriving the weights wi,t (as e.g. following the traditional Markowitz

approach), we follow Brandt et al. (2009) and parameterize the weights as a function of firm

characteristics xi,t , i.e.

wi,t = f ( xi,t ; θ ), (2)

where θ is the coefficient vector to be estimated.

The parameter vector θ remains constant across assets i and periods t, i.e. it maximizes the

conditional expected utility at every period t. This necessarily implies that θ also maximizes

the unconditional expected utility. Hence, one can estimate θ by maximizing the unconditional

expected utility via the return distribution’s sample analogues:

!
T T Nt
1 1
∑ u r p,t+1 (θ ) = T ∑ u ∑ f (xi,t ; θ )ri,t+1


max . (3)
θ T t =1 t =1 i =1

The idea behind parametric portfolio policies is that one may exploit firm characteristics in

order to tilt some benchmark portfolio towards stocks that increase an investor’s utility, so that

f (·) can be expressed as

1
wi,t = bi,t + g( xi,t ; θ ), (4)
Nt

where bi,t denotes benchmark portfolio weights such as the equally weighted or value weighted

7
portfolio and x̂i,t denotes the characteristics of stock i, standardized cross-sectionally to have zero

mean and unit standard deviation in each cross section t.1

Brandt et al. (2009) and the subsequent literature (e.g. DeMiguel et al., 2020) restrict firm

characteristics to affect the portfolio in a linear, additive manner, such that

1 T
wi,t = bi,t + θ x̂i,t . (5)
Nt

In essence, our model can be interpreted as a generalization of the linear parametric portfolio

policy approach, as we allow x̂i,t to enter the model flexibly and non-linearly. More specifically, we

allow g(·) in equation (4) to take arguably one of the most flexible forms - a feed-forward neural

network. As discussed in the introduction, this represents a significant conceptual devivation

from the literature in at least two respects: First, by replacing the linear specification with a neural

network, we allow the relationship between firm characteristics and weights to be non-linear, and

we account for potential interactions of firm characteristics, in line with the recent literature that

finds that such flexiblity can be important to predict expected return (Moritz and Zimmermann,

2016; Freyberger et al., 2020; Gu et al., 2020). Here, our approach explores whether such flexibility

also helps to model the relationship between portfolio weights and firm characteristics. Second, this

flexibility comes at the cost of having to estimate a model with a high-dimensional parameter

vector. Thus, it departs from the original motivation of the parametric portfolio policy literature,

which aimed to reduce portfolio optimization to a low-dimensional problem where only a small

number of coefficients need to be estimated. Our benchmark model has about 5,700 parameters

compared to the three parameters that need to be estimated when using Brandt et al. (2009).

Why might non-linear modeling of portfolio weights be important? Consider an investor

who trades off mean return against return volatility.The investor uses standard one-dimensional

portfolio sorting techniques as pictured in Figure C.1 in Appendix C. Decile portfolios formed

on short-term reversal or sales-to-price display monotonically increasing mean return.2 At the

same time, the standard deviations of decile portfolios are non-linear in deciles, in particular
1 The 1/N term is a normalization that allows the portfolio weight function to be applied to a time-varying number
t
of stocks. Without this normalization, an increase in the number of stocks with an otherwise unchanged cross-sectional
distribution of characteristics leads to more radical allocations, although the investment opportunities are basically
unchanged.
2 We picked these two variables for illustrative purposes as these variables are the most important return- and

fundamental-based variables in Gu et al. (2020).

8
top and bottom decile portfolios display high standard deviation. This leads to the extreme

portfolios having comparatively low Sharpe ratios relative to decile portfolios in the middle of

the distribution. A mean-variance (long-only) investor would therefore potentially be indifferent

between investing in any portfolio in the upper half of the short-term reversal distribution, and

she would prefer to invest in portfolios in the middle of the sales-to-price distribution rather than

investing in the extreme portfolios. It is these kinds of relationships that a non-linear portfolio

policy can capture. On top of modeling such non-linearities, our models below also allow for

interactions between different signal variables that cannot be represented by one-dimensional

portfolio sorts either.

2.2 Network architecture

We implement and compare a range of so-called feed-forward networks, a popular network

structure that is prominently used in prediction contexts such as image recognition but has also

recently been applied to stock return prediction. Conceptually, our feed-forward networks are

structured to estimate optimal portfolio weights and as such differ from networks used in pure

prediction contexts in two important ways.

First, the objective of our estimation is to maximize expected utility. Standard use of predictive

modeling (with or without networks) tries to minimize some distance metric (e.g. mean squared

error) between e.g. observed stock returns and predicted stock returns. For example, Gu et al.

(2020) use neural networks to predict stock returns using a penalized mean squared error as the

statistical loss function.

In contrast, we follow Brandt et al. (2009) and directly estimate portfolio weights. More

specifically, we predict portfolio weights by maximizing the unconditional sample analogue of a

utility function as given in equation (3). For example, in our base case, the loss function L that we

aim to minimize with respect to θ is the negative standard mean-variance utility:

 !2 
T T
1 γ 1
L(θ ) =
T ∑ 2 r p,t+1 (θ ) −
T ∑ r p,t+1 (θ ) − r p,t+1 (θ ) , (6)
t =1 t =1

where γ is the absolute risk aversion parameter. Note that minimizing Equation (6) is equivalent

9
to maximizing mean-variance utility.

Second, our loss function requires the portfolio return per period t, so that we need to aggregate

our outputs cross-sectionally in each period. To do so, we maintain the three-dimensional structure

of our data, i.e. we do not treat it as two-dimensional as e.g. Gu et al. (2020) do. Conceptually,

our models can be depicted as shown in Figure 1.

[FIGURE 1 ABOUT HERE]

In Figure 1, the input data on the left form a cube (or 3D tensor) with dimensions time t,

stocks i and input variables k. Input data are fed into networks with different numbers of hidden

layers.3 In line with equation (4), the output of the neural network is then normalized by 1/Nt

and added to the benchmark portfolio b. The output of the model O is a two-dimensional matrix

with dimensions t × i of portfolio weights for each stock and time period.

Constructing a neural network requires many design choices, including the depth (number

of layers) and width (units per layer) of the model, respectively. Recent literature suggests that

deeper networks can achieve higher accuracy with less width than wider models (Eldan and

Shamir, 2016). However, for smaller data sets a large number of parameters can lead to overfitting

and/or issues in regards to the optimization process. Selecting the best network structure is a

formidable task and not our main objective.4 Instead, we rely on the results of Gu et al. (2020) and

use their most successful model as our benchmark model. We explore robustness of our findings

to changes in both network complexity and network structure in Appendix B.

As discussed in Section 2.1, the network’s output needs to be normalized and can be interpreted

as the deviation from a benchmark portfolio. In our application, the benchmark portfolio is the

equally weighted portfolio in all models. A common alternative would be a value weighted

benchmark portfolio where weights are determined by a stock’s market capitalization. We stick

to the equally weighted benchmark because of empirical evidence that it outperforms other

benchmarks like the value weighted benchmark for longer periods (DeMiguel et al., 2009).

Lastly, we control for unreasonable results and overfitting in terms of portfolio weights by
3 Following Feng et al. (2018) and Bianchi et al. (2020) we only count the number of hidden layers while excluding

the output layer in the remainder of this paper.

4 In practice, the task is often approximated by comparing a few different structures and selecting the one with the

best performance.

10
ex-ante imposing an upper bound on an individual stock’s absolute portfolio weight of |3%|, i.e.

|wi,t | ≤ 0.03. (7)

In doing so, we ensure that the model performance does not rely too heavily on particular stocks.

We employ a range of different additional regularization techniques that are standard in the deep

learning literature. We give an outline of these techniques and a more detailed description of the

structure of the model including its hyperparameters in Appendix A.

2.3 Data

We use the Open Source Asset Pricing dataset of Chen and Zimmermann (2022). The dataset

contains monthly US stock-level data on 205 cross-sectional stock return predictors, covering the

period from January 1925 to December 2020.

We focus on the period from January 1971 to December 2020, since comprehensive accounting

data is only sparsely available in the years prior to that. In addition, we also only keep common

stocks, i.e. stocks with share codes 10 and 11, and stocks that are traded on the NYSE (exchange

code equal to 1) to ensure that results are not driven by small stocks. We match the data with

monthly stock return data from the Center for Research in Security Prices (CRSP). We drop any

observation with missing return, size and/or a return of less than -100%. We include continuous

firm characteristics from Chen and Zimmermann (2022)’s categories Price, Trading, Accounting and

Analyst, respectively.5

Finally, we follow Gu et al. (2020) and replace missing values with the cross-sectional median

at each month for each stock, respectively. Additionally, similar to Gu et al. (2020) we rank all

stock characteristics cross-sectionally. As in Brandt et al. (2009) and DeMiguel et al. (2020), each

predictor is then standardized to have a cross-sectional mean of zero and standard deviation of

one. Note that each predictor is signed so that a larger value implies a higher expected return.

Our final dataset contains 157 predictors for a total of 5,154 firms. Each month, the dataset
5 All characteristics are calculated at a monthly frequency. For variables that are updated at a lower frequency, the

monthly value is simply the last observed value. We assume the standard lag of six months for annual accounting
data availability and a lag of one quarter for quarterly accounting data availability. For IBES, we assume that earnings
estimates are available by the end date of the statistical period. For other data, we follow the respective original research
in regards to availability.

11
contains a minimum of 1,213, a maximum of 1,855 and an average of 1,422 firms. Table D.1 in the

appendix lists the included predictors by original paper. The three columns in the table describe

the update frequency of each predictor, the predictor category and the economic category, both

taken from Chen and Zimmermann (2022). As part of our robustness check, we exploit that

information in Appendix B to construct non-fully connected networks.

2.4 Out-of-sample testing strategy

Following Brandt et al. (2009) and Gu et al. (2020), we use an expanding window strategy to

generate out-of-sample results. More specifically, we split our data into a training sample used to

estimate the model, a validation sample used to tune the hyperparameters of the model and a test

sample used to evaluate the out-of-sample performance of the model.

We initially train the model on the first 20 years of the dataset, validate it on the following five

years and evaluate its out of-sample-performance on the year following the validation window.

We then recursively increase the training sample by one year. Each time the training sample is

increased, we refit the entire model while holding the size of the validation and test window fixed.

The result is a sequence of out-of-sample periods corresponding to each expanding window, in

our case 25 in total. Note that this approach ensures that the temporal ordering of the data is

maintained. The testing strategy is depicted graphically in Figure 2.

[FIGURE 2 ABOUT HERE]

2.5 Model interpretation

Machine learning models are notoriously difficult to interpret and neural networks are no

exception. Nevertheless, in our application, understanding the estimated relation between input

(firm characteristics) and output (estimated portfolio weights) is essential in order to shed light on

the relation between firm characteristics and utility. Moreover, such an understanding allows us

to compare our results to the existing literature. We provide three ways of interpreting the models

and of identifying the most important predictors among the plethora of variables that enter our

models.

12
 First, we evaluate the extent to which non-linearity in variables (i.e. variable interactions) and

non-linearity in parameters (i.e. functional form) contribute to the estimated deep parametric

portfolio policy. Put differently, we assess the extent to which different forms of non-linearity

play a role when optimizing portfolios conditional on firm characteristics. To do so, we estimate

a linear surrogate model in which we regress the out-of-sample weight predictions on all firm

characteristics. This allows us to assess the extent to which a simple linear model is capable of

ex-post explaining the predicted weights. In a next step, we estimate a second surrogate model,

this time including all possible two-way interactions, i.e. allowing for non-linearity in variables.

To prevent excessive overfitting, we also add a lasso penalty term. This allows us to assess to

which extent non-linearity in variables play a role in regards to predicting weights. We attribute

the remaining unexplained portion of predicted deep parametric portfolio weights to the effect

of non-linearity in functional form. Furthermore, we analyze the portfolio characteristics of the

ex-post fitted surrogate models during the out of sample periods. Inter alia, this enables us

to assess to which extent non-linearity with respect to weight predictions translates into utility

differences.

Second, we calculate variable importance in the model as the decrease in model performance

when a particular variable is missing from the model. That is, for every out-of-sample period we

set all values of a variable to zero while holding the remaining variables fixed. We then calculate

the utility loss as compared to the original model in every out-of-sample period and take the

average across all models. For the sake of comparability, we scale the average utility losses across

all variables for each model so that they add up to one. As a result, we are able to rank the

variables according to the average utility loss that occurs if they are excluded from the model.

Third, we evaluate the sensitivity of the model output to each variable. Typically, partial

dependence plots provide an assessment of the variables of interest over a range of values. At each

value of the variable, the model is evaluated while the remaining variables remain unchanged,

and the results are then averaged across the cross-section. However, since the sum of all weights

in each cross-section is equal to one and thus the mean weight prediction is always the same,

applying this method to parametric portfolio policies does not yield reasonable results. To

circumvent this problem, we apply our own algorithm: when assessing the sensitivity with respect

to variable k, we set the values of the remaining variables to zero, i.e. their median. This means

13
that effectively, we reduce our input data to the variable of interest. We then predict out-of-sample

portfolio weights based on the estimated model and the manipulated data. Subsequently, we

plot the weights as a function of input variable k. We interpret the behavior of predicted weights

conditional on values of k as the sensitivity of weights (i.e. its partial dependence) with respect to

k.

3 Results

3.1 Benchmark portfolios

Table 1 presents the comparison between different portfolios based on their utility, weights and

return characteristics. We compare a simple equally weighted and a value weighted portfolio

with the parametric portfolio policy of Brandt et al. (2009) and our own deep parametric portfolio

policy.6 Analogous to Brandt et al. (2009) we provide results as follows: We report (1) the utility

that a respective portfolio strategy generates, (2) distributional characteristics of the portfolio

weights, (3) properties of the portfolio returns and (4) the strategies’ alphas against a Fama-French

six-factor model.

The first row of Table 1 reports the realized utility across out-of-sample periods for a mean-

variance investor with absolute risk aversion of five. The equally weighted and value weighted

portfolio yield a utility of 0.0024 and 0.0029, respectively. The standard parametric portfolio policy

substantially outperforms the simple portfolios, yielding a utility of 0.0267. However, the deep

parametric portfolio policy yields a utility of 0.0469, almost twice as large as the utility derived

from the linear parametric portfolio policy. The difference in utilities is significant at the 0.1%

level.7 This suggests that taking into account predictor interactions and non-linear relationships

substantially improves an investor’s utility.

The next set of rows gives insight into the distribution of the respective portfolio weights. The

active portfolios take comparably large positions, with the average absolute weight of the deep
6 Toensure comparability between the linear and deep parametric portfolio policy we differ slightly from Brandt
et al. (2009) in that the linear model includes l1 -regularization and early stopping, similar to the deep model. A more
detailed description is given in Appendix A.
7 We follow DeMiguel et al. (2022) and construct one-sided p-values from 10,000 bootstrap samples using the

stationary bootstrap method of Politis and Romano (1994) with an average block size of five and the procedure of
Ledoit and Wolf (2008). This method is also used when assessing the statistical significance of utility and Sharpe ratio
differences between the deep and the linear parametric portfolio policy hereafter.

14
portfolio policy being almost nine times as large as in the case of the equally weighted and value

weighted portfolio, respectively. However, due to the weight constraint shown in Equation (7)

these positions remain below 3%. Although the average absolute weight is larger in the deep

model as compared to the linear model, the maximum (1.7% versus 2.1%) and minimum weights

(-1.8% versus -2.2%) are smaller. Comparing the actively managed portfolios, we find that both

have similar levels of leverage, with the deep parametric policy being slightly higher (387% versus
+
315%), yet producing almost twice as much turnover (770% versus 394%), where wi,t −1 is the

portfolio before rebalancing at time t, that is,

+
wi,t −1 = wi,t−1 ∗ (1 + ri,t ). (8)

As Ang et al. (2011) show, average gross leverage of hedge fund companies amounts to 120% in

the period after the financial crisis 2007-2008. This indicates that both the linear and the deep

portfolio policies are rather unrealistic in the benchmark case. We address this in Section 4.2 by

including a penalty term for turnover and a constraint for leverage in our objective function.

The monthly mean returns of 4.7% and 7% in the linear and deep policy case are much higher

than the mean returns of around 1.1% in the equally weighted and value weighted portfolio

cases due to their highly levered nature. Note that our deep model yields a 2.3 percentage

point increase as compared to the linear policy, while its standard deviation increases only

modestly by 0.7 percentage points, thereby leading to a Sharpe ratio that is around 40% higher.

The difference in Sharpe ratios is statistically significant at the 1% level. In fact, both models

substantially outperform the market porfolios with more than twice as large Sharpe ratios. In

terms of skewness and kurtosis the deep portfolio policy stands out as compared to the other

portfolios. In particular, the portfolio exhibits a positive skewness (1.05) and high kurtosis (6.51).

However, the third and fourth moments are of no interest to an investor with mean-variance

preference.

The bottom set of rows reports the alphas and its standard errors with respect to a six-factor

model that appends a momentum factor to the Fama-French five-factor model. The market

portfolio alphas are both not significantly different from zero. The linear policy alpha is 3.2%.

The deep policy alpha is even higher, amounting to 5.6%. Both alphas are highly statistically

15
significant. These large unexplained returns can partially be attributed to the highly levered

nature of the active portfolios, as we show in the following sections.

[TABLE 1 ABOUT HERE]

These results are robust to changing the network architecture as we show in Appendix B.

More specifically, we confirm our results for different levels of model complexity and non-fully

connected networks.

3.2 Surrogate model, variable importance and partial dependence

Surrogate model

Surrogate modeling allows us to disentangle the contributions of non-linearity in variables

and non-linearity in functional form with respect to the predictions as well as the utility gains of

the deep parametric portfolio policy as compared to the linear parametric portfolio policy. As

one would expect, Figure 3 shows that a simple linear surrogate model perfectly explains the

out-of-sample weight predictions retrieved from the linear parametric portfolio policy. However, a

simple linear model only explains 60-70% of the variation in out-of-sample weights predicted by

the deep parametric portfolio policy. An extended surrogate model that allows for non-linearity

in variables explains between 80-88% of the variation in out-of-sample weights. Based on these

numbers, one can infer that up to ∼70% of the underlying characteristic-weight relationship is

of linear nature, ∼10-20% can be captured by interactions, and the remaining ∼10-20% can be

captured by a non-linear functional form.

[FIGURE 3 ABOUT HERE]

In Table 2 we further analyze the portfolios generated by the the respective surrogate models.

As implicitly indicated by the aforementioned R2 ’s, the surrogate model for the parametric

portfolio policy yields a portfolio that is equivalent to the original model. Hence, its portfolio

characteristics, especially its utility, are equivalent to the original model. In the deep parametric

portfolio policy case, three observations stand out: First, the simple linear surrogate model yields

16
a utility that is nearly 20% lower than that of the original deep parametric portfolio policy. Second,

the linear surrogate model extended by two-way interactions yields a utility that is slightly

higher than the utility based on the simple linear surrogate model. Thus, the above mentioned

observations in terms of R2 roughly translate into differences in utility. However, this difference is

not as large as the difference in terms of R2 . Third, note that the utility of a linear surrogate model

which is fitted to the estimated portfolio weights of the deep parametric portfolio policy (DPPP

Pred.) is much larger than the utility of the linear portfolio policy that is directly fitted to the data

(the original PPP model). This can seem puzzling at first sight as the PPP model should be able to

generate the same weights. The reason is that surrogate models are fitted on the out-of-sample

weight predictions. Hence, the predictions of the linear surrogate model yield in-sample utility,

while the original models show out-of-sample utility.

[TABLE 2 ABOUT HERE]

Variable importance

Next, we turn to variable importance measured as discussed in section 2.5. Figure 4 compares

the most important variables in the linear and deep parametric portfolio policies. For both

models, we find that the majority of the most important predictors relate to past returns. Short-

term reversal is the most important variable in both models, mirroring findings in Moritz and

Zimmermann (2016) and Gu et al. (2020). The deep parametric portfolio policy is even more

tilted towards such variables. In particular, out of the twenty most important variables in the

linear parametric portfolio case, eleven are price-related, seven are accounting-related and two

are analyst-related. In the deep parametric portfolio case, fourteen of the twenty most important

variables are price-related, five are accounting-related and one is analyst related. As past-return

based variables typically imply higher turnover, this is consistent with the higher turnover of the

resulting portfolio policy.

[FIGURE 4 ABOUT HERE]

Figure C.2 in the appendix attempts to group variables into categories (such as "earnings-

related", or "risk-related"). We do again find that the most important categories contain past-return

17
based variables. The other relevant variable categories are "delayed processing" (anomalies that

are based on delayed processing of information, e.g. industry momentum) and earnings-related

variables.8

Partial dependence

Figure 5 depicts the marginal association between portfolio weights and input variables. We

examine the sensitivity with respect to three fundamental variables, namely the book-to-market

ratio (BM), liquid assets (cash), and quarterly return on assets (roaq), as well as an analyst variable,

namely earnings forecast revisions per share (AnalystRevision), and four past return-based

variables, namely 12-month momentum (Mom12m), short-term reversal (STreversal), seasonal

momentum (MomSeason), and intermediate momentum (IntMom). Recall that each predictor

is signed, so that a larger value implies a higher expected return. As implied by the linearity of

the approach, the variables are linearly related to predicted weights in the case of the standard

linear parametric portfolio policy. In contrast, the deep parametric portfolio policy weights are

non-linearly related to the variables. More specifically, these relationships all appear to be convex.

Interestingly, the convex shape appears to be quite similar for every variable: a steep increase

in weight prediction occurs in the sixth or seventh decile, respectively. Moreover, the weight

predictions generally appear to roughly follow the trend in mean returns across deciles. The

difference in marginal sensitivities between the linear and the deep parametric portfolio policy

illustrates that the latter is picking up non-linear relationships that the former is not able to pick

up by construction.

[FIGURE 5 ABOUT HERE]

4 Extensions of the benchmark model

4.1 Long only

A large majority of equity portfolios face restrictions on short selling. We incorporate short-sale

constraints as in Brandt et al. (2009), i.e. we truncate portfolios weights at zero (and still keep the
8 Table D.1 in the appendix shows the category of each anomaly variable, based on Jensen et al. (2021) and extended

by us for variables that are not considered in their study.

18
cap of 3% per stock). In particular, to make sure that portfolio weights still sum up to one, we add

the following portfolio rebalancing term to the end of our optimization process:

∗ max [0, wi,t ]

wi,t = Nt
. (9)
∑ max [0, wi,t ]
j =1

Table 3, shows results from estimating long-only portfolios. Again, the deep parametric portfo-

lio policy yields the highest utility, although utility is markedly lower than in the unconstrained

case. Still, the utility of the deep parametric portfolio policy is around four times higher than the

utility of the market portfolios and around 40% higher than the utility of the linear parametric

portfolio policy. The difference between the utility of the deep and the linear parametric portfolio

policy is statistically significant at the 0.1% level.

Both active portfolios result in a much higher turnover than the market portfolios, and the

deep portfolio policy produces a higher turnover than the linear portfolio policy (125% versus

72%). Different from the unconstrained benchmark results in Table 1, here we report the fraction

of weights that are equal to zero. Interestingly, on average the deep portfolio policy does not

include 11% of stocks, while the linear portfolio policy does not include 32% of the available

stocks. Thus, the deep portfolio policy invests in more stocks but also has a higher individual

maximum weight (1.64% vs 0.42%), indicating that many weights are possibly very low.

The deep portfolio policy yields higher expected returns than the linear portfolio policy, with

a moderate increase in volatility resulting in a Sharpe ratio that is around 19% higher than the

Sharpe ratio of the linear portfolio policy. This difference is statistically significant at the 0.1%

level. Interestingly, the third and fourth moments of all portfolio policies are similar and the

portfolio return distributions are not heavily skewed or tailed. Lastly, the alphas of the Fama-

French model are a lot smaller, while still being highly significant in both the linear and the

deep portfolio policy case. Without the ability to take (potentially extreme) short positions, the

estimated parametric portfolios appear to be much more realistic. Nonetheless, the deep portfolio

policy still outperforms the other portfolios in terms of realized out-of-sample utility.

The comparison between the unconstrained (Table 1) and the long-only case (Table 3) also

yields interesting insights. First, the unconstrained portfolio benefits from using the short positions

19
as leverage to increase exposure to the long positions. Consistent with this observation, the linear

portfolio policy has a similar fraction of short positions and stocks not held in the two models.

Second, the maximum weight of the linear portfolio policy decreases by around 80% in the

long-only case as compared to the unconstrained case. Interestingly, both findings do not apply to

the deep portfolio policy. The fraction of short positions is a lot higher than the fraction of stocks

not held in the long-only deep portfolio policy. Moreover, the maximum weight is similar in the

unconstrained and constrained case. This can be attributed to the non-linearity of the deep model.

Variable importance rankings are similar to the unconstrained models. Figure 6 shows the

variable importance of the 50 most important firm characteristics, ranked by average importance

across all models. These include the two benchmark models, the linear and deep long-only models,

and the linear and deep constraint models from Section 4.2. Each column corresponds to a single

model, and the color gradations within each column indicate the most important (black) to least

important (white) firm characteristics. The third and fourth columns correspond to the long-only

models and show that the importance of the variables is similar to the benchmark models. In

both the unconstrained and the long only models, characteristics based on past returns are at

the top, with short-term reversal being the most important variable in three of the four models.

In the linear long-only model the industry return of big firms (IndRetBig) exhibits the highest

importance. Moreover, the importance in terms of values is similar between the benchmark and

the long-only models. To conclude, these results show that the long-only investor also relies

heavily on past return-based characteristics.

[FIGURE 6 ABOUT HERE]

4.2 Transaction costs and leverage

The results of the unconstrained linear and the deep portfolio policy yield unfeasible portfolios

with high leverage and turnover. To investigate whether the deep portfolio policy also outperforms

the regular portfolio policy in a more realistic setting, we include a penalty term for transactions

costs similar to DeMiguel et al. (2020) and include an additional constraint for maximum leverage.

In our estimation, we use estimated transaction costs from Chen and Velikov (2021).9 Thus,
9 We thank the authors for making an updated version of the data available.

20
analogously, we define transaction costs κi,t as the effective half bid-ask spread. We follow

DeMiguel et al. (2020) in constructing the penalty term added to the policy optimization as

Nt
TC = Et [ ∑ |κi,t (wi,t − wi,t
+
−1 )|], (10)
i =1

+
where wi,t −1 is the portfolio before rebalancing as in Equation (8).

The leverage constraint is constructed analogously to our weight constraint in Equation (7).

Ang et al. (2011) show that the average gross leverage of hedge fund companies amounts to 120%

in the period after the financial crisis 2007-2008. We use a slightly more conservative number of a

maximum leverage of 100%. The penalty is constructed such that the gross leverage cannot exceed

100% in a single period in model training. This constraint is formulated for every period t as

Nt
∑ wi I ( wi < 0) ≥ −1 (11)
i =1

for each period, where I (wi < 0) is a vector where an element is one if the corresponding portfolio

weight is smaller than zero and zero otherwise.

Table 4 shows the results for the constrained optimization process. We see that the constraints

lead to a decrease in utility for the deep and linear policy. The utility decrease is greater

for the deep portfolio policy. Both estimated portfolios still outperform the market portfolios.

Interestingly, the constraints lead to the deep portfolio policy being much closer to the linear one.

This indicates that the deep model exploits the short-selling ability and characteristics with high

turnover more extensively than the linear model. More specifically, the deep model predicts high

weights in good performing stocks at the cost of less diversification. Still, the deep parametric

portfolio policy delivers a utility gain over a linear policy of about 20%, statistically significant at

the 1% level. Thus, despite turnover still being higher than in the linear approach (168% versus

97%), the deep model still yields a higher realized mean-variance utility. Overall, in both models,

the maximum and minimum positions are less extreme than in the unconstrained case and thus

more realistic compared to the unconstrained case.

Furthermore, mean return and variance decrease in both active models. However, the linear

portfolio policy only suffers a small decrease in Sharpe ratio, while the deep portfolio policy’s

21
Sharpe ratio decreases by around a third. Nonetheless, the difference between Sharpe ratios is

still significant at the 5% level. The third and fourth moment are similar across all portfolios. The

alphas of the estimated models are much smaller, but still highly significant.

[TABLE 4 ABOUT HERE]

Comparing the variable importance of included firm characteristics with the previous models,

we find that this set of constraints leads to a very different picture. Figure 6 shows the importance

of the variables for the constrained models in columns five and six. The figure illustrates that

the importance of characteristics based on past returns is much lower compared to the previous

four models. Overall, short-term reversal loses its place as the most important variable in the

linear model. This is an intuitive result, since trading conditional on short-term reversal implies

turnover by definition. Hence, when penalizing turnover via transaction costs, short-term reversal

will necessarily lose some of its importance to a certain degree. Further, in line with the results

of DeMiguel et al. (2020), we observe that variable importance is much more balanced across

variables in general. In the deep model, short-term reversal is still the most important variable,

but it becomes evident that its relative importance is lower than in the previous models. Again,

this intuitively follows from the aforementioned mechanism. As in the linear model, variable

importance becomes more balanced in the deep model when introducing transaction costs and

leverage constraints. This is also underlined by lower (higher) maximum (minimum) portfolio

weights compared to the previous models. The mean absolute portfolio weights are also much

smaller than for the benchmark portfolios. This shows that the constraints lead to a more

diversified portfolio, which is reflected in a more balanced importance of firm characteristics.

5 Different investor utility functions

5.1 Different risk aversion parameters

Different investors may exhibit different levels of risk aversion. In our benchmark model we

assume an absolute risk aversion coefficient of five. Table 5 shows how our model performs for

different degrees of absolute risk aversion in the mean-variance case. In order to meaningfully

22
interpret the differences in utility, we do not report utility itself, but rather the difference in utility

relative to a constant benchmark, i.e. an equally weighted portfolio. Other than that, we report

the same result metrics as before.

The results show that investors with an absolute risk aversion of five experience the largest

utility gains relative to the equally weighted portfolio benchmark. In general, we observe that the

utility gains decrease relative to an equally weighted portfolio with higher risk aversion, which is

due to the fact that the portfolio of the highly risk averse investor is more diversified and therefore

closer to the equally weighted portfolio. Further, this shows that the risk aversion parameter

can also be used as a regularization parameter, since increasing risk aversion leads to decreasing

variance in the predicted weights, which reduces overfitting. Consequently, the investor with

a risk aversion of two achieves lower utility gains relative to the benchmark portfolio than the

investor with a risk aversion of five due to overfitting.

We further observe a negative correlation between risk aversion and absolute portfolio weights

as well as leverage and turnover. This aligns with the intuition of more risk averse investors not

focusing on single high return characteristics, but rather on diversifying their portfolio with a

more balanced weight distribution. This in turn results in portfolios that display lower expected

returns, but also lower volatility for more risk averse investors. Moreover, all portfolios seem to

have a similar Sharpe ratios. The third and fourth moment of the portfolio return distributions

tend to be less extreme the higher the risk aversion, indicating that the higher the risk aversion, the

more the respective portfolio return distribution tends towards a normal distribution. Intuitively,

with increasing risk aversion the alphas of the factor model regressions decrease.

[TABLE 5 ABOUT HERE]

5.2 CRRA and loss aversion

Analogously to varying risk aversion for a mean-variance investor, we can account for different

investor types by changing the utility function in our optimization process in Equation (1). In

particular, we explore linear and deep portfolio policies for an investor with constant relative risk

23
aversion utility defined as
(1 + r p,t+1 )1−γ
u(r p,t+1 ) = , (12)
1−γ

where γ is the relative risk aversion of the investor, and for a loss-averse investor (Tversky and

Kahneman (1992)) with utility defined as


b

 − l (W − ( 1 + r

p,t+1 )) if (1 + r p,t+1 ) < W
u(r p,t+1 ) = , (13)
((1 + r p,t+1 ) − W )b

 otherwise

where W is a reference wealth level determined in the editing stage, the parameter l measures the

investor‘s loss aversion and the parameter b captures the degree of risk seeking over losses and

risk aversion over gains.

Table 6 reports the results for the linear and deep portfolio policy for an investor with constant

relative risk aversion of five and an investor with a subjective wealth level W equal to one, loss

aversion of 2.5 and parameter value b equal to one which corresponds to pure loss aversion.

Interestingly, for both preferences the deep portfolio policy achieves higher utility than the linear

portfolio policy.

The results for the CRRA preferences are similar to those for mean-variance preferences with

similar risk aversion, except that the third and fourth moment of the deep policy are not as

extreme. The differences in the higher moments can be attributed to the investor’s preference

over higher order moments, which differentiate the CRRA investor from an equally risk-averse

mean-variance investor. In our data, however, the effect of higher order moments is not strong

enough to heavily change the portfolio weight distribution and the resulting portfolio returns.

By far the most interesting part of the loss averse investor’s preference is the fact that she cares

about the size of the tail of the portfolio return distribution, rather than the mean to variance ratio,

which is relevant to a mean-variance investor. This is also reflected in the results in Figure 6. Both

portfolios show a high variance compared to the mean-variance and CRRA investor, however,

they also display higher skewness. This high positive skewness illustrates a highly right tailed

distribution. As the p-value indicates, the Sharpe ratios do not seem to differ significantly, while

the deep model results in more than twice the utility than the linear model. The deep portfolio

24
policy yields a high variance paired with a high skewness and kurtosis. Thus, the portfolio return

distribution is heavily tailed to the right with no particularly high losses. The weight distribution

of the portfolios is still very similar to the other utility models, while the deep portfolio policy

yields slightly higher leverage and turnover.

[TABLE 6 ABOUT HERE]

We further investigate how utility is accumulated across out-of-sample periods. More specif-

ically, for each utility function, we plot the cumulative out-of-sample period per period utility

of the equally weighted portfolio, the linear parametric portfolio and the deep parametric port-

folio. Irrespective of the investor’s utility function, the deep parametric portfolio consistently

outperforms the linear parametric portfolio policy and an equally weighted portfolio in utility

terms.

[FIGURE 7 ABOUT HERE]

6 Conclusion

Building on the parametric portfolio policy of Brandt et al. (2009), we show that feed-forward

neural networks can be used to optimize portfolios based on a large number of firm characteristics

for different investor preferences. We develop a flexible framework that can be used to implement

neural networks for portfolio choice problems to optimize different utility functions with flexible

constraints. More specifically, we show that neural networks can be used to optimize portfolio

weighting based on firm characteristics in a one-step optimization framework. Furthermore, we

show how traditional distance loss functions can be replaced by context-specific utility functions

in neural networks.

Our empirical results indicate that neural networks perform significantly better than linear

models in regards to portfolio allocation, suggesting that firm characteristics are non-linearily

related to optimal portfolio weights. Consistent with this hypothesis, we show that linear surrogate

models are not able to fully explain the deep parametric portfolio weight predictions, even when

25
accounting for two-way interactions. We further shed light on the non-linear relationship between

characteristics and predicted weights by depicting the sensitivity of predicted weights with respect

to the input. Again, we show a clearly non-linear effect. Gaining more insights into the models,

we find that return-based stock characteristics resemble the most important group of predictors.

However, consistent with DeMiguel et al. (2020), variable importance is more evenly distributed

and puts less weight on past returns when constraints and transaction costs are taken into account

in the optimization process.

Exploring variations in the degree of an investor’s risk aversion or their utility function, we

find that (absolute) portfolio weights are typically lower when risk aversion is higher, consistent

with more risk-averse investors aiming for a more balanced portfolio weight distribution.

Overall, the results show that neural networks are successful in solving portfolio choice

problems. Specifically, this is due to neural networks allowing predictor variables to relate to

moments of the expected return distribution non-linearly, both in terms of variable interactions

and in terms of functional form. Highlighting the growing role of machine learning and non-

linear models in finance, our approach thus resembles a comparably simple and flexible neural

network based model that enables practitioners and researchers alike to create reasonable portfolio

allocations based on firm characteristics and preferences.

26
References

Ammann, M., G. Coqueret, and J.-P. Schade (2016). Characteristics-based portfolio choice with

leverage constraints. Journal of Banking & Finance 70, 23–37.

Ang, A., S. Gorovyy, and G. B. van Inwegen (2011). Hedge fund leverage. Journal of Financial

Economics 102(1), 102–126.

Bianchi, D., M. Büchner, and A. Tamoni (2020). Bond Risk Premiums with Machine Learning. The

Review of Financial Studies 34(2), 1046–1089.

Brandt, M. W., P. Santa-Clara, and R. Valkanov (2009). Parametric Portfolio Policies: Exploiting

Characteristics in the Cross-Section of Equity Returns. The Review of Financial Studies 22(9),

3411–3447.

Butler, A. and R. H. Kwon (2021). Integrating prediction in mean-variance portfolio optimization.

Working Paper.

Chen, A. Y. and M. Velikov (2021). Zeroing in on the Expected Returns of Anomalies. Working

Paper.

Chen, A. Y. and T. Zimmermann (2022). Open source cross-sectional asset pricing. Critical Finance

Review 27(2), 207–264.

Chevalier, G., G. Coqueret, and T. Raffinot (2022). Supervised portfolios. Quantitative Finance 22(12),

2275–2295.

Cong, L., K. Tang, J. Wang, and Y. Zhan (2021). Alphaportfolio: Direct construction through deep

reinforcement learning and interpretable ai. Working Paper.

DeMiguel, V., L. Garlappi, and R. Uppal (2009). Optimal Versus Naive Diversification: How

Inefficient is the 1/N Portfolio Strategy? The Review of Financial Studies 22(5), 1915–1953.

DeMiguel, V., A. Martin-Utrera, and R. Uppal (2022). A multifactor perspective on volatility-

managed portfolios. Working Paper.

27
DeMiguel, V., A. Martín-Utrera, F. J. Nogales, and R. Uppal (2020). A Transaction-Cost Perspective

on the Multitude of Firm Characteristics. The Review of Financial Studies 33(5), 2180–2222.

Eldan, R. and O. Shamir (2016). The power of depth for feedforward neural networks. In

V. Feldman, A. Rakhlin, and O. Shamir (Eds.), 29th Annual Conference on Learning Theory,

Volume 49 of Proceedings of Machine Learning Research, pp. 907–940.

Fama, E. F. and K. R. French (2008). Dissecting anomalies. The Journal of Finance 63(4), 1653–1678.

Farrell, M. H., T. Liang, and S. Misra (2021). Deep learning for individual heterogeneity: An

automatic inference framework. Working Paper.

Feng, G., J. He, and N. G. Polson (2018). Deep Learning for Predicting Asset Returns. Working

Paper.

Freyberger, J., A. Neuhierl, and M. Weber (2020). Dissecting characteristics nonparametrically. The

Review of Financial Studies 33(5), 2326–2377.

Goodfellow, I., Y. Bengio, and A. Courville (2016). Deep Learning. MIT Press.

Green, J., J. R. M. Hand, and X. F. Zhang (2017, 03). The Characteristics that Provide Independent

Information about Average U.S. Monthly Stock Returns. The Review of Financial Studies 30(12),

4389–4436.

Gu, S., B. Kelly, and D. Xiu (2020). Empirical Asset Pricing via Machine Learning. The Review of

Financial Studies 33(5), 2223–2273.

Hansen, L. and P. Salamon (1990). Neural network ensembles. IEEE Transactions on Pattern Analysis

and Machine Intelligence 12(10), 993–1001.

Heaton, J. B., N. G. Polson, and J. H. Witte (2017). Deep learning for finance: deep portfolios.

Applied Stochastic Models in Business and Industry 33(1), 3–12.

Hjalmarsson, E. and P. Manchev (2012). Characteristic-based mean-variance portfolio choice.

Journal of Banking & Finance 36(5), 1392–1401.

28
Ioffe, S. and C. Szegedy (2015, 07–09 Jul). Batch normalization: Accelerating deep network training

by reducing internal covariate shift. Proceedings of the 32nd International Conference on Machine

Learning 37, 448–456.

Jensen, T. I., B. T. Kelly, S. Malamud, and L. H. Pedersen (2022). Machine learning and the

implementable efficient frontier. Swiss Finance Institute Research Paper No. 22-63.

Jensen, T. I., B. T. Kelly, and L. H. Pedersen (2021). Is there a replication crisis in finance? Technical

report, National Bureau of Economic Research.

Kelly, B. T., S. Malamud, and K. Zhou (2022). The virtue of complexity in machine learning

portfolios. Swiss Finance Institute Research Paper Series (21-90).

Kingma, D. P. and J. Ba (2014). Adam: A method for stochastic optimization. Working Paper.

Ledoit, O. and M. Wolf (2008). Robust performance hypothesis testing with the sharpe ratio.

Journal of Empirical Finance 15(5), 850–859.

Markowitz, H. (1952). Portfolio selection. The Journal of Finance 7(1), 77–91.

Masters, T. (1993). Practical Neural Network Recipes in C++. Academic Press Professional, Inc.

Moritz, B. and T. Zimmermann (2016). Tree-based conditional portfolio sorts: The relation between

past and future stock returns. Working Paper.

Novy-Marx, R. and M. Velikov (2016). A taxonomy of anomalies and their trading costs. The

Review of Financial Studies 29(1), 104–147.

Politis, D. N. and J. P. Romano (1994). The stationary bootstrap. Journal of the American Statistical

Association 89(428), 1303–1313.

Srivastava, N., G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov (2014). Dropout:

A simple way to prevent neural networks from overfitting. Journal of Machine Learning Re-

search 15(56), 1929–1958.

Tversky, A. and D. Kahneman (1992). Advances in prospect theory: Cumulative representation of

uncertainty. Journal of Risk and Uncertainty 5(4), 297–323.

29
Uysal, A. S., X. Li, and J. M. Mulvey (2021). End-to-end risk budgeting portfolio optimization

with neural networks. Working Paper.

30
Figures

Input Hidden Output

layer layer layer

I1 H1
b

I2
..
i data
.. .
O
i weights

t
k
. Hm
t
Ik
1/Nt

Figure 1: Neural Network Structure

This figure presents the core structure of our neural networks. White circles denote the input layer, grey
circles denote the hidden layer and black circles denote the output layer. The data cube on the left depicts
the structure of our data, i.e. we have k variables across i cross-sections in t periods. The rectangle on the
right depicts our output, i.e. weights across i cross-sections in t periods. The output of the neural network
is normalized by 1/Nt and added to the benchmark portfolio b. The final output is labeled O.

31
 periods
0 5 10 15 20 25 30 35 40 45 50

2
Training
windows

3 Validation
Test
4

...

30

Figure 2: Out-Of-Sample Testing Strategy

This figure presents our out-of-sample testing strategy. We recursively increase our training window,
presented by the black portion of each bar, while holding validation and test window constant, presented
by the grey portions of each bar.

32
 1.0

0.9
R squared

0.8

0.7

0.6

1995 2000 2005 2010 2015 2020

Years

Model PPP DPPP DPPP^2

Figure 3: Surrogate R2
This figure depicts the R2 of the surrogate models in the benchmark case. More specifically, the "PPP"-line
depicts the R2 of a linear surrogate model in case of the PPP, the "DPPP"-line depicts the R2 of a linear
surrogate model in case of the DPPP and the "DPPP2 "-line depicts the R2 of a l 1 -regularized linear surrogate
model including first order effects and all possible two-way interactions.

33
 STreversal
IndRetBig
AnnouncementReturn
MomSeason
EntMult
IntMom
Mom12m
MomSeasonShort
ChTax
cfp
AnalystRevision
CF
MomSeason06YrPlus
Cash
CBOperProf
EarningsSurprise
ResidualMomentum
EarningsStreak
Coskewness
MaxRet
0.00 0.05 0.10 0.15
Importance

(a) PPP

STreversal
IndRetBig
AnnouncementReturn
MomSeason
IntMom
High52
AnalystRevision
LRreversal
IndMom
BMdec
Mom12m
MRreversal
MaxRet
MomSeason06YrPlus
roaq
MomOffSeason06YrPlus
Cash
MomSeasonShort
EntMult
EarningsStreak
0.00 0.05 0.10 0.15
Importance

(b) DPPP

Figure 4: Variable importance for PPP and DPPP

Variable importance for the 20 most influential variables in the linear and deep parametric portfolio policy.
Variable importance is an average over all training samples and normalized to sum to one.

34
 BM Cash roaq AnalystRevision
0.020
0.006
0.015
0.004
0.010
0.005 0.002
Mean Return

Weights
0.000 0.000
STreversal Mom12m MomSeason IntMom
0.020
0.006
0.015
0.004
0.010
0.005 0.002

0.000 0.000
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Deciles

Model PPP DPPP

Figure 5: Marginal association between portfolio weights and characteristics

This figure shows the sensitivity of predicted weights (right vertical axis) with respect to values of the
respective variable (horizontal axis). The aforementioned relationship is depicted by curves, smoothed via
spline-regressions. The figure also includes bars, depicting the mean return (left vertical axis), per variable
decile (horizontal axis).

35
 STreversal
IndRetBig
AnnouncementReturn
MomSeason
IntMom
AnalystRevision
Mom12m
MomOffSeason06YrPlus
EntMult
roaq
ChTax
MomSeason06YrPlus
High52
MomSeasonShort
Coskewness
EarningsStreak
CF
LRreversal
Cash
EarningsSurprise
MomSeason11YrPlus
MaxRet
MRreversal
MomSeason16YrPlus
Mom12mOffSeason
cfp
IdioVolAHT
CBOperProf
Beta
BM
CompEquIss
NumEarnIncrease
retConglomerate
ResidualMomentum
IndMom
RD
IdioVol3F
AbnormalAccruals
NetEquityFinance
Price
IntanBM
RDcap
VolumeTrend
Mom6m
ChNNCOA
BMdec
zerotradeAlt12
IntanCFP
IdioRisk
OperProfRD
PPP_Main DPPP_Main PPP_Long DPPP_Long PPP_Con DPPP_Con

Figure 6: Variable importance across models

We rank the top 50 stock characteristics in terms of its importance across all models. The higher a stock
characteristic within the figure, the higher its average importance across all models. Columns correspond to
individual models, with columns ending with "_Main" representing unconstrained models, columns ending
with "_Long" representing long-only models, and columns ending with "_Con" representing models with
constrained leverage and transaction costs. The color gradations within each column indicate importance,
i.e. the darker the gradation, the more important the stock characteristic.

36
 Mean-variance

10

0
2000 2010 2020
Cummulative utility

CRRA
0
-20
-40
-60

2000 2010 2020

Loss aversion

15
10
5
0
-5
2000 2010 2020
Date

Portfolio EW PPP DPPP

Figure 7: Cumulative utility for different utility functions

This figure includes plots of the cumulative utility for each of the utility functions considered. More
specifically, for each utility function, we plot the cumulative utility across out-of-sample periods for the
equally weighted portfolio (EW), the parametric portfolio policy (PPP) and the deep parametric portfolio
policy (DPPP).

37
Tables

Table 1: Deep and linear portfolio policy

EW VW PPP DPPP
Utility 0.0024 0.0029 0.0267 0.0469
p-value(UDPPP − UPPP ) 0.0004

∑ |wi |/Nt ∗ 100 0.0694 0.0694 0.5060 0.6057

max wi ∗ 100 0.0704 0.1113 2.0748 1.7260
min wi ∗ 100 0.0704 0.0410 -2.2097 -1.8370
∑ wi I ( wi < 0) 0.0000 0.0000 -3.1475 -3.8665
∑ I (wi < 0)/Nt 0.0000 0.0000 0.4334 0.4411
+
∑ |wi,t − wi,t −1 | 0.0931 0.0779 3.9370 7.6984

Mean 0.0110 0.0105 0.0468 0.0701

StdDev 0.0587 0.0552 0.0897 0.0965
Skew -0.3716 -0.5039 -0.1451 1.0537
Kurt 3.6591 3.3455 1.8391 6.5084
SR 0.6461 0.6609 1.8070 2.5170
p-value(SR DPPP − SR PPP ) 0.0066

FF5 + Mom α -0.0002 -0.0003 0.0323 0.0559

StdErr (α) 0.0007 0.0006 0.0040 0.0051

This table shows out-of-sample estimates of the deep and linear portfolio policies optimized for a mean-
variance investor with absolute risk aversion of five conditional on 157 firm characteristics. The regular
portfolio policy is a linear specification of Equation (4), while the deep model is a feed-forward neural
network with three hidden layers and 32, 16, and eight nodes, respectively. We use data from the Open
Source Asset Pricing Dataset from January 1971 to December 2020. The columns labeled "EW", "VW", "PPP"
and "DPPP" show the statistics of the equally-weighted portfolio, value-weighted portfolio, parametric
portfolio policy, and deep parametric portfolio policy, respectively. The first rows show the utility of the
investor as well as the bootstrapped one-sided p-value for the difference in utility between the DPPP and
the PPP. The second set of rows shows statistics on portfolio weights averaged over time. These statistics
include the average absolute portfolio weight, the average maximum and minimum portfolio weights,
the average sum of negative weights in the portfolio, the average proportion of negative weights in the
portfolio, and the turnover in the portfolio. The third set of rows shows the first four moments of the
final portfolio return distributions as well as the annualized Sharpe ratios and the bootstrapped one-sided
p-value for the difference in Sharpe ratios between the DPPP and the PPP. The bottom panel shows the
alphas and their standard errors with respect to the Fama-French five-factor model extended to include the
momentum factor.

38
Table 2: Surrogate models

PPP PPP Pred. DPPP DPPP Pred. DPPPˆ2 Pred.

R2 0.9999 0.6607 0.8552

Utility 0.0267 0.0266 0.0469 0.0382 0.0398

∑ |wi |/Nt ∗ 100 0.5060 0.5059 0.6057 0.4674 0.4619

max wi ∗ 100 2.0748 2.0744 1.7260 2.0096 1.9105
min wi ∗ 100 -2.2097 -2.2095 -1.8370 -2.0663 -2.2662
∑ wi I ( wi < 0) -3.1475 -3.1474 -3.8665 -2.8697 -2.8300
∑ I (wi < 0)/Nt 0.4334 0.4335 0.4411 0.4412 0.4290
+
∑ |wi,t − wi,t −1 | 3.9370 3.9365 7.6984 6.2002 6.1322

Mean 0.0451 0.0450 0.0684 0.0537 0.0563

StdDev 0.0895 0.0896 0.0963 0.0831 0.0854
Skew -0.1636 -0.2002 1.0761 0.2137 0.3930
Kurt 1.9043 2.0386 6.6301 2.6537 2.8152
SR 1.7449 1.7403 2.4591 2.2388 2.2840

This table shows out-of-sample estimates of the deep and linear portfolio policies with 157 firm char-
acteristics and optimized for a mean-variance investor with absolute risk aversion of five as well as the
estimates of the linear surrogate models for the PPP and the DPPP, respectively. The regular portfolio
policy is a linear specification of Equation (4), while the deep model is a feed-forward neural network
with three hidden layers and 32, 16, and eight nodes, respectively. The columns labeled "PPP" and "DPPP"
show the statistics of the originally estimated portfolio policies, while the columns labeled "PPP Pred."
and "DPPP Pred." show the statistics of the linear surrogate model. Finally, the column labeled "DPPPˆ2
Pred." shows the statistics of a lasso surrogate model that includes the predictors and all possible two-way
interactions. The first row shows the average R2 of regressing out-of-sample weight predictions on the
respective surrogate model. The second row shows the utility of the investor. The third set of rows shows
statistics on portfolio weights averaged over time. These statistics include the average absolute portfolio
weight, the average maximum and minimum portfolio weights, the average sum of negative weights in the
portfolio, the average proportion of negative weights in the portfolio, and the turnover in the portfolio. The
fourth set of rows shows the first four moments of the final portfolio return distributions as well as the
annualized Sharpe ratios.

39
Table 3: Long-only deep and linear portfolio policy

EW VW PPP DPPP
Utility 0.0024 0.0029 0.0084 0.0116
p-value(UDPPP − UPPP ) 0.0001

∑ |wi |/Nt ∗ 100 0.0694 0.0694 0.0694 0.0694

max wi ∗ 100 0.0704 0.1113 0.4155 1.6420
min wi ∗ 100 0.0704 0.0410 0.0000 0.0000
∑ wi I ( wi < 0) 0.0000 0.0000 0.0000 0.0000
∑ I (wi < 0)/Nt 0.0000 0.0000 0.3173 0.1148
+
∑ |wi,t − wi,t −1 | 0.0931 0.0779 0.7222 1.2519

Mean 0.0110 0.0105 0.0153 0.0198

StdDev 0.0587 0.0552 0.0526 0.0573
Skew -0.3716 -0.5039 -0.5551 -0.4191
Kurt 3.6591 3.3455 3.5843 4.0876
SR 0.6461 0.6609 1.0045 1.1941
p-value(SR DPPP − SR PPP ) 0.0002

FF5 + Mom α -0.0002 -0.0003 0.0048 0.0090

StdErr (α) 0.0007 0.0006 0.0008 0.0011

This table shows out-of-sample estimates of the deep and linear portfolio policies including a long-only
constraint optimized for a mean-variance investor with absolute risk aversion of five conditional on 157 firm
characteristics. The regular portfolio policy is a linear specification of Equation (4), while the deep model
is a feed-forward neural network with three hidden layers and 32, 16, and eight nodes, respectively. We
use data from the Open Source Asset Pricing Dataset from January 1971 to December 2020. The columns
labeled "EW", "VW", "PPP" and "DPPP" show the statistics of the equally-weighted portfolio, value-weighted
portfolio, parametric portfolio policy, and deep parametric portfolio policy, respectively. The first rows show
the utility of the investor as well as the bootstrapped one-sided p-value for the difference in utility between
the DPPP and the PPP. The second set of rows shows statistics on portfolio weights averaged over time.
These statistics include the average absolute portfolio weight, the average maximum and minimum portfolio
weights, the average sum of negative weights in the portfolio, the average proportion of negative weights in
the portfolio, and the turnover in the portfolio. The third set of rows shows the first four moments of the
final portfolio return distributions as well as the annualized Sharpe ratios and the bootstrapped one-sided
p-value for the difference in Sharpe ratios between the DPPP and the PPP. The bottom panel shows the
alphas and their standard errors with respect to the Fama-French five-factor model extended to include the
momentum factor.

40
Table 4: Constrained and penalized deep and linear portfolio policy

EW VW PPP DPPP
Utility 0.0021 0.0028 0.0139 0.0169
p-value(UDPPP − UPPP ) 0.0015

∑ |wi |/Nt ∗ 100 0.0694 0.0694 0.1749 0.1819

max wi ∗ 100 0.0704 0.1113 0.6827 0.7866
min wi ∗ 100 0.0704 0.0410 -0.6817 -0.9814
∑ wi I ( wi < 0) 0.0000 0.0000 -0.7607 -0.8113
∑ I (wi < 0)/Nt 0.0000 0.0000 0.3417 0.3181
+
∑ |wi,t − wi,t −1 | 0.0931 0.0779 0.9699 1.6756

Mean 0.0107 0.0104 0.0184 0.0223

StdDev 0.0584 0.0549 0.0421 0.0465
Skew -0.3711 -0.5055 -0.9085 -0.7414
Kurt 3.6640 3.3435 2.6099 2.8353
SR 0.6345 0.6533 1.5110 1.6624
p-value(SR DPPP − SR PPP ) 0.0451

FF5 + Mom α -0.0003 -0.0004 0.0076 0.0113

StdErr (α) 0.0007 0.0006 0.0013 0.0017

This table shows out-of-sample estimates of the deep and linear portfolio policies including a transaction
cost penalty and a leverage constraint, optimized for a mean-variance investor with absolute risk aversion
of five conditional on 157 firm characteristics. The regular portfolio policy is a linear specification of
Equation (4), while the deep model is a feed-forward neural network with three hidden layers and 32,
16, and eight nodes, respectively. We use data from the Open Source Asset Pricing Dataset from January
1971 to December 2020. The columns labeled "EW", "VW", "PPP" and "DPPP" show the statistics of the
equally-weighted portfolio, value-weighted portfolio, parametric portfolio policy, and deep parametric
portfolio policy, respectively. The first rows show the utility of the investor as well as the bootstrapped
one-sided p-value for the difference in utility between the DPPP and the PPP. The second set of rows shows
statistics on portfolio weights averaged over time. These statistics include the average absolute portfolio
weight, the average maximum and minimum portfolio weights, the average sum of negative weights in the
portfolio, the average proportion of negative weights in the portfolio, and the turnover in the portfolio. The
third set of rows shows the first four moments of the final portfolio return distributions (net of transaction
costs) as well as the annualized Sharpe ratios and the bootstrapped one-sided p-value for the difference in
Sharpe ratios between the DPPP and the PPP. The bottom panel shows the alphas and their standard errors
with respect to the Fama-French five-factor model extended to include the momentum factor.

41
Table 5: Deep portfolio policy for mean-variance investors with different degrees of risk aversion

γ=2 γ=5 γ = 10 γ = 20
% Utility Increase 780.4002 1885.2435 565.3362 122.6475

∑ |wi |/Nt ∗ 100 0.6749 0.6057 0.5295 0.3847

max wi ∗ 100 1.8125 1.7260 1.6331 1.2971
min wi ∗ 100 -1.8523 -1.8370 -1.8039 -1.3872
∑ wi I ( wi < 0) -4.3656 -3.8665 -3.3171 -2.2737
∑ wi I (wi < 0)/Nt 0.4451 0.4411 0.4344 0.4171
+
∑ |wi,t − wi,t −1 | 8.5704 7.6984 6.7283 4.8273

Mean 0.0786 0.0701 0.0628 0.0482

StdDev 0.1115 0.0965 0.0824 0.0656
Skew 1.3035 1.0537 0.3598 0.5061
Kurt 8.2253 6.5084 0.9416 1.3940
SR 2.4408 2.5170 2.6402 2.5443

FF5 + Mom α 0.0626 0.0559 0.0492 0.0368

StdErr (α) 0.0058 0.0051 0.0043 0.0033

This table shows out-of-sample estimates of the deep portfolio policies, optimized for a mean-variance
investor with absolute risk aversion of two, five, ten and 20, respectively, conditional on 157 characteristics.
The deep model is a feed-forward neural network with three hidden layers and 32, 16, and eight nodes,
respectively. We use data from the Open Source Asset Pricing Dataset from January 1971 to December 2020.
The columns labeled "γ = 2", "γ = 5", "γ = 10" and "γ = 20" show the statistics of the deep parametric
portfolio policy with risk aversion of two, five, ten and 20, respectively. The first row shows the difference
in utility relative to an equally weighted portfolio. The second set of rows shows statistics on portfolio
weights averaged over time. These statistics include the average absolute portfolio weight, the average
maximum and minimum portfolio weights, the average sum of negative weights in the portfolio, the
average proportion of negative weights in the portfolio, and the turnover in the portfolio. The third set
of rows shows the first four moments of the final portfolio return distributions as well as the annualized
Sharpe ratios. The bottom panel shows the alphas and their standard errors with respect to the Fama-French
five-factor model extended to include the momentum factor.

42
Table 6: Deep portfolio policy with different investor preferences

CRRA LA
PPP DPPP PPP DPPP
Utility -0.2253 -0.2063 0.0266 0.0574
p-value(UDPPP − UPPP ) 0.0003 0.0004

∑ |wi |/Nt ∗ 100 0.4972 0.6127 0.5034 0.6468

max wi ∗ 100 2.0363 1.7452 2.0743 1.7618
min wi ∗ 100 -2.1712 -1.8709 -2.1577 -1.7841
∑ wi I ( wi < 0) -3.0841 -3.9171 -3.1290 -4.1627
∑ I (wi < 0)/Nt 0.4351 0.4430 0.4307 0.4490
+
∑ |wi,t − wi,t −1 | 3.7816 7.8053 3.7464 8.3677

Mean 0.0473 0.0711 0.0473 0.0783

StdDev 0.0890 0.0982 0.0871 0.1359
Skew -0.1004 0.8169 0.0996 3.5153
Kurt 1.3766 4.9609 0.8451 33.2542
SR 1.8391 2.5101 1.8789 1.9963
p-value(SR DPPP − SR PPP ) 0.0075 0.4227

FF5 + Mom α 0.0324 0.0570 0.0338 0.0624

StdErr (α) 0.0040 0.0052 0.0040 0.0067

This table shows out-of-sample estimates of the deep and linear portfolio policies, optimized for an investor
with constant relative risk aversion preference (CRRA) with relative risk aversion of five and a loss averse
(LA) investor with loss aversion of 2.5, subjective wealth level of one and degree of risk seeking of one,
respectively, conditional on 157 firm characteristics. The regular portfolio policy is a linear specification of
Equation (4), while the deep model is a feed-forward neural network with three hidden layers and 32, 16,
and eight nodes, respectively. We use data from the Open Source Asset Pricing Dataset from January 1971
to December 2020. The columns labeled "PPP" and "DPPP" show the statistics of the parametric portfolio
policy, and deep parametric portfolio policy, respectively. The first rows show the utility of the investor
as well as the bootstrapped one-sided p-value for the difference in utility between the DPPP and the PPP.
The second set of rows shows statistics on portfolio weights averaged over time. These statistics include
the average absolute portfolio weight, the average maximum and minimum portfolio weights, the average
sum of negative weights in the portfolio, the average proportion of negative weights in the portfolio, and
the turnover in the portfolio. The third set of rows shows the first four moments of the final portfolio
return distributions as well as the annualized Sharpe ratios and the bootstrapped one-sided p-value for the
difference in Sharpe ratios between the DPPP and the PPP. The bottom panel shows the alphas and their
standard errors with respect to the Fama-French five-factor model extended to include the momentum
factor.

43
Description of appendices

• Appendix A: Neural Network Configuration

• Appendix B: Robustness

• Appendix C: Supplementary figures

• Appendix D: Supplementary tables

44
Appendix A Neural Network Configuration

Our benchmark model consists of an input layer, three hidden layers and an output layer. We

apply the geometric pyramid rule (Masters, 1993), i.e. the first hidden layer consists of 32 nodes,

the second hidden layer consists of 16 nodes and the third hidden layer consists of eight nodes.

We consider different network architectures in Appendix B.

At each node of the network, a linear transformation of the preceding outputs is fed into

an activation function. We choose to use the leaky rectified linear unit (leaky ReLU) activation

function at every node. 


z
 if z > 0
R(z) = , (14)


αz otherwise

where z denotes the input and α denotes some small non-zero constant, in our case 0.01. ReLU is

the most popular activation function because it is cheap to compute, converges fast and is sparsely

activated. The disadvantage of transforming all negative values to zero is a problem called "dying

ReLU". A ReLU neuron is "dead" if it is stuck in the negative range and always outputs zero. Since

the slope of ReLU in the negative range is also zero, it is unlikely that a neuron will recover once

it goes negative. Such neurons play no role in discriminating inputs and are essentially useless.

Over time, a large part of the network may do nothing. Leaky ReLU fixes this problem because

it has small slope for negative values instead of a flat slope. Moreover, we shift the activation

function at every node in every hidden layer by adding a constant. This is commonly referred to

as bias in the machine learning literature.

Our benchmark network is estimated by minimizing the loss function (utility function) given

in Equation (6). To do so, we apply the commonly used ADAM stochastic gradient descent

optimization technique developed by Kingma and Ba (2014).

To control for the non-linearity and heavy parametrization of the model, we employ different

regularization techniques to prevent overfitting: first, as mentioned above, we impose a constraint

on an individual stock’s absolute portfolio weight of |3%|.

Second, we add a lasso (l1 ) penalty term to the loss function to be minimized. Adding the

penalty implies a potential shrinkage of coefficients towards 0. This in turn reduces the variance

of the prediction, i.e. preventing the model to be overfitted.

45
 Third, we employ early stopping on the validation data. Early stopping refers to a very general

regularization technique. At each new iteration, predictions are estimated for the validation

sample, and the loss (utility) is constructed. The optimization is terminated when the validation

sample loss starts to increase by some small specified number (tolerance) over a specified number

of iterations (patience). Typically, the termination occurs before the loss is minimized in the training

sample. Early stopping is a popular regularization tool because it reduces the computational cost.

Fourth, we implement a dropout layer before the first hidden layer (Srivastava et al., 2014).

The basic idea of dropout is to randomly remove units (and their connections) from the neural

network during training. This prevents the units from becoming too similar. During training,

samples are taken from an exponential number of different thinned networks. At test time, it

is easy to approximate the effect of averaging the predictions of all these thinned networks by

simply using a single, unthinned network with smaller weights. The combination of a dropout

layer, l1 -regularization and early stopping tremendously helps to reduce overfitting and model

complexity.

Fifth, we adopt an ensemble approach in training our neural network (Hansen and Salamon,

1990). In particular, we initialize five neural networks with different random seeds and construct

predictions by averaging the predictions from all networks. This reduces the variance across

predictions since different seeds produce different predictions due to the stochastic nature of the

optimization process.

Finally, we adopt our own version of a batch normalization algorithm (Ioffe and Szegedy,

2015). In general, training deep neural networks is complicated by the fact that the distribution of

inputs to each layer changes during training as the parameters of the previous layers change. This

phenomenon is referred to as internal covariate shift and can be remedied by normalizing the layer

inputs. The strength of this method is that normalization is part of the model architecture and is

performed for each training mini-batch. Batch normalization allows much higher learning rates to

be used and less care to be taken in initialization. Brandt et al. (2009) standardize characteristics

cross-sectionally to have zero mean and unit standard deviation across all stocks at date t. Hence,

the model predictions represent deviations from the benchmark portfolio. However, applying the

aforementioned activation function destroys this structure. In our model each observation can be

interpreted as a complete cross-section (e.g. a batch size of 12 refers to 12 complete cross-sections

46
of data). However, the model of Brandt et al. (2009) requires normalization on a cross-sectional

level instead of a batch level. Thus, we employ our own version of cross-sectional normalization

after applying the activation function in each hidden layer, such that the output of each node in

the hidden layer is standardized cross-sectionally to have zero mean and unit standard deviation

across all stocks at date t. Hence, the output of each node in each hidden layer can also be

interpreted as a deviation from the benchmark portfolio.

We provide a summary of the relevant hyperparameters in Table D.2.

[TABLE D.2 ABOUT HERE]

47
Appendix B Robustness Checks

B.1 Model complexity

Our benchmark model is a relatively shallow neural net with only three hidden layers. It is

conceivable that a more complex model can achieve even higher utility gains over a linear model.

For example, Goodfellow et al. (2016) observe that neural nets with more hidden layers tend to

outperform neural nets with fewer hidden layers but more nodes per layer. Kelly et al. (2022)

report evidence in support of complex models in the context of forecasting aggregate stock market

returns.

We extend our benchmark model to include between two and five hidden layers. All models

start with 32 nodes in the first hidden layer and then halve the number of nodes in each subse-

quent layer. The number of parameters across models therefore varies between 5,600 and 5,768.

Additionally, we add different possible learning rates to our hyperparameter tuning and increase

the number of epochs and patience for early stopping, to account for the different complexities of

the models and to ensure that more complex models also reach their respective potential.

Table D.3 shows the results. The second model is our original benchmark model that we

added for comparison.10 The remaining columns contain results based on networks with two,

four or five hidden layers. We observe that reducing the number of hidden layers to two slightly

reduces the utility. This reduction in utility is significant at the 10%-level. In contrast, increasing

the number of hidden layers to four or five, respectively, does not yield statistically significant

differences in utility. We thus conclude that in general, reasonable complexity adjustments in

terms of the number of hidden layers do not lead to significantly different outcomes.

[TABLE D.3 ABOUT HERE]

B.2 Non-fully connected networks

Theoretically, there is a large range of different options to how one may adjust the network

structure. In this section, we explore one structural change. Following Bianchi et al. (2020), we
10 Notethat the utility slightly differs from our benchmark in Section 3.1. This is due to the aforementioned fact that
we add different possible learning rates as well as increase the number of epochs and patience for early stopping. We
do so not only for the model variations, but also for our benchmark to ensure consistency across models.

48
split our input according to its characteristics and feed the resulting input groups separately into

the model. This is illustrated in Figure C.3.

[Figure C.3 ABOUT HERE]

More specifically, we split our data according to its update frequency and its data category,

respectively. For update frequency we divide our data into monthly, yearly and quarterly

characteristics. For data category we divide our data into Accounting, Price, Trading and Analyst

characteristics. The update frequency and data category of each predictor is shown in Table D.1 in

the Appendix.

We interact only characteristics with the same frequency (category) in the first hidden layer

which can be interpreted as a dimension reduction for each frequency (category). After that we

proceed with the ordinary network architecture in the second and third hidden layer. These are

just two different network structure variations out of the plethora of different possibilities.

Table D.4 shows the results for the benchmark linear and deep portfolio policy followed by

the two variations in network architectures for the deep portfolio policy. The results indicate

that changes in realized utility are not large. In fact, splitting according the predictor category

does not yield significant gains or losses in terms of utility. Splitting according to the frequency

of predictors does lead to a small increase of utility, which is significant at the 10%-level. Both

new models produce slightly higher leverage and turnover than the base deep portfolio policy.

Moreover, the new models yield higher Sharpe ratios by reducing the variance of the portfolio

return distributions. The largest differences can be observed for the third and fourth moment of

the return distribution, where both new models show less extreme skewness and kurtosis which

results in more realistic return distributions.

[TABLE D.4 ABOUT HERE]

49
Appendix C Supplementary Figures

50
 0.015
0.015

Mean return

Mean return
0.010
0.010

0.005
0.005

0.000 0.000
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

0.20
0.20

0.15
0.15
Return SD

Return SD
0.10 0.10

0.05 0.05

0.00 0.00
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

0.12
51

0.09 0.09
Sharpe ratio

Sharpe ratio
0.06 0.06

0.03 0.03

0.00 0.00
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Quantile of STreversal Quantile of SP

(a) Short-Term Reversal (b) Sales-To-Price

Figure C.1: Mean returns, standard deviations and Sharpe ratios of one-dimensional portfolio sorts
Mean returns, standard deviations and Sharpe ratios of decile portfolios sorted on short-term reversal (left panel) and sales-to-price ratio (right
panel).
 short-term reversal
delayed processing
momentum
earnings
tax
volatility
long term reversal
risk
valuation
profitability
asset composition
external financing
R&D
recommendation
liquidity
sales growth
accruals
volume
other
leverage
investment
investment growth
cash flow risk
composite accounting
size
PPP_Main DPPP_Main PPP_Long DPPP_Long PPP_Con DPPP_Con

Figure C.2: Variable importance per cluster

We group the variables into clusters according to the economic category specified in the Open Source Asset
Pricing data set by Chen and Zimmermann (2022). Clusters are then ranked by average characteristic
importance within the respective cluster. The higher a cluster within the figure, the higher its average
importance across all models. Columns correspond to individual models, with columns ending with
"_Main" representing unconstrained models, columns ending with "_Long" representing long-only models,
and columns ending with "_Con" representing models with constrained leverage and transaction costs. The
color gradations within each column indicate importance, i.e. the darker the gradation, the more important
the cluster.

52
 Input Hidden Output
layer layer layer

I1 H1
b

I2
..
i data
.. .
O
i weights

t
k
. Hm
t
Ik
1/Nt

Figure C.3: Non-Fully Connected Neural Network Structure

This figure presents the structure of our non-fully connected networks. White circles denote the input layer,
grey circles denote the hidden layer and black circles denote the output layer. The data cube on the left
depicts the structure of our data, i.e. we have k variables across i cross-sections in t periods. The rectangle
on the right depicts our output, i.e. weights across i cross-sections in t periods. The output of the neural
network is normalized by 1/Nt and added to the benchmark portfolio b. The final output is labeled O.

53
Appendix D Supplementary Tables

54
 Acronym Long Description Author(s) Year, Journal Frequency Cat.Data Cat.Economic

ChInvIA Change in capital inv (ind adj) Abarbanell and Bushee 1998, AR yearly Accounting investment growth
GrSaleToGrInv Sales growth over inventory growth Abarbanell and Bushee 1998, AR yearly Accounting sales growth
GrSaleToGrOverhead Sales growth over overhead growth Abarbanell and Bushee 1998, AR yearly Accounting sales growth
IdioVolAHT Idiosyncratic risk (AHT) Ali, Hwang, and Trombley 2003, JFE monthly Price volatility
EarningsConsistency Earnings consistency Alwathainani 2009, BAR yearly Accounting earnings
Illiquidity Amihud’s illiquidity Amihud 2002, JFM monthly Trading liquidity
BidAskSpread Bid-ask spread Amihud and Mendelsohn 1986, JFE monthly Trading liquidity
grcapx Change in capex (two years) Anderson and Garcia-Feijoo 2006, JF yearly Accounting investment growth
grcapx3y Change in capex (three years) Anderson and Garcia-Feijoo 2006, JF yearly Accounting investment growth
betaVIX Systematic volatility Ang et al. 2006, JF monthly Price volatility
IdioRisk Idiosyncratic risk Ang et al. 2006, JF monthly Price volatility
IdioVol3F Idiosyncratic risk (3 factor) Ang et al. 2006, JF monthly Price volatility
55

CoskewACX Coskewness using daily returns Ang, Chen and Xing 2006, RFS monthly Price risk
Mom6mJunk Junk Stock Momentum Avramov et al 2007, JF monthly Price momentum
OrderBacklogChg Change in order backlog Baik and Ahn 2007, Other yearly Accounting accruals
roaq Return on assets (qtrly) Balakrishnan, Bartov and Faurel 2010, JAE quarterly Accounting profitability
MaxRet Maximum return over month Bali, Cakici, and Whitelaw 2010, JF monthly Price volatility
ReturnSkew Return skewness Bali, Engle and Murray 2015, Book monthly Price risk
ReturnSkew3F Idiosyncratic skewness (3F model) Bali, Engle and Murray 2015, Book monthly Price risk
CBOperProf Cash-based operating profitability Ball et al. 2016, JFE yearly Accounting profitability
OperProfRD Operating profitability R&D adjusted Ball et al. 2016, JFE yearly Accounting profitability
Size Size Banz 1981, JFE monthly Price size
SP Sales-to-price Barbee, Mukherji and Raines 1996, FAJ yearly Accounting valuation
EP Earnings-to-Price Ratio Basu 1977, JF monthly Price valuation

Continued on next page

 Acronym Long Description Author(s) Year, Journal Frequency Cat.Data Cat.Economic

InvGrowth Inventory Growth Belo and Lin 2012, RFS yearly Accounting profitability
BrandInvest Brand capital investment Belo, Lin and Vitorino 2014, RED yearly Accounting investment
Leverage Market leverage Bhandari 1988, JFE monthly Price leverage
ResidualMomentum Momentum based on FF3 residuals Blitz, Huij and Martens 2011, JEmpFin monthly Price momentum
Price Price Blume and Husic 1972, JF monthly Price other
NetPayoutYield Net Payout Yield Boudoukh et al. 2007, JF monthly Price valuation
PayoutYield Payout Yield Boudoukh et al. 2007, JF monthly Price valuation
NetDebtFinance Net debt financing Bradshaw, Richardson, Sloan 2006, JAE yearly Accounting external financing
NetEquityFinance Net equity financing Bradshaw, Richardson, Sloan 2006, JAE yearly Accounting external financing
XFIN Net external financing Bradshaw, Richardson, Sloan 2006, JAE yearly Accounting external financing
DolVol Past trading volume Brennan, Chordia, Subra 1998, JFE monthly Trading volume
FEPS Analyst earnings per share Cen, Wei, and Zhang 2006, WP monthly Analyst profitability
56

AnnouncementReturn Earnings announcement return Chan, Jegadeesh and Lakonishok 1996, JF monthly Price earnings
REV6 Earnings forecast revisions Chan, Jegadeesh and Lakonishok 1996, JF monthly Analyst earnings
AdExp Advertising Expense Chan, Lakonishok and Sougiannis 2001, JF monthly Accounting R&D
RD R&D over market cap Chan, Lakonishok and Sougiannis 2001, JF monthly Accounting R&D
CashProd Cash Productivity Chandrashekar and Rao 2009, WP yearly Accounting profitability
std_turn Share turnover volatility Chordia, Subra, Anshuman 2001, JFE monthly Trading liquidity
VolSD Volume Variance Chordia, Subra, Anshuman 2001, JFE monthly Trading liquidity
retConglomerate Conglomerate return Cohen and Lou 2012, JFE monthly Price delayed processing
RDAbility R&D ability Cohen, Diether and Malloy 2013, RFS yearly Accounting other
AssetGrowth Asset growth Cooper, Gulen and Schill 2008, JF yearly Accounting investment
EarningsForecastDisparity Long-vs-short EPS forecasts Da and Warachka 2011, JFE monthly Analyst earnings
CompEquIss Composite equity issuance Daniel and Titman 2006, JF monthly Accounting external financing

Continued on next page

 Acronym Long Description Author(s) Year, Journal Frequency Cat.Data Cat.Economic

IntanBM Intangible return using BM Daniel and Titman 2006, JF yearly Accounting long term reversal
IntanCFP Intangible return using CFtoP Daniel and Titman 2006, JF yearly Accounting long term reversal
IntanEP Intangible return using EP Daniel and Titman 2006, JF yearly Accounting long term reversal
IntanSP Intangible return using Sale2P Daniel and Titman 2006, JF yearly Accounting long term reversal
ShareIss5Y Share issuance (5 year) Daniel and Titman 2006, JF monthly Accounting external financing
LRreversal Long-run reversal De Bondt and Thaler 1985, JF monthly Price long term reversal
MRreversal Medium-run reversal De Bondt and Thaler 1985, JF monthly Price long term reversal
EquityDuration Equity Duration Dechow, Sloan and Soliman 2004, RAS yearly Price valuation
cfp Operating Cash flows to price Desai, Rajgopal, Venkatachalam 2004, AR yearly Accounting valuation
ForecastDispersion EPS Forecast Dispersion Diether, Malloy and Scherbina 2002, JF monthly Analyst volatility
ExclExp Excluded Expenses Doyle, Lundholm and Soliman 2003, RAS quarterly Analyst composite accounting
ProbInformedTrading Probability of Informed Trading Easley, Hvidkjaer and O’Hara 2002, JF yearly Trading liquidity
57

OrgCap Organizational capital Eisfeldt and Papanikolaou 2013, JF yearly Accounting R&D
sfe Earnings Forecast to price Elgers, Lo and Pfeiffer 2001, AR monthly Analyst valuation
GrLTNOA Growth in long term operating assets Fairfield, Whisenant and Yohn 2003, AR yearly Accounting investment
AM Total assets to market Fama and French 1992, JF yearly Accounting valuation
BMdec Book to market using December ME Fama and French 1992, JPM yearly Accounting valuation
BookLeverage Book leverage (annual) Fama and French 1992, JF yearly Accounting leverage
OperProf operating profits / book equity Fama and French 2006, JFE yearly Accounting profitability
Beta CAPM beta Fama and MacBeth 1973, JPE monthly Price risk
EarningsSurprise Earnings Surprise Foster, Olsen and Shevlin 1984, AR quarterly Analyst earnings
AnalystValue Analyst Value Frankel and Lee 1998, JAE monthly Analyst valuation
AOP Analyst Optimism Frankel and Lee 1998, JAE monthly Analyst other
PredictedFE Predicted Analyst forecast error Frankel and Lee 1998, JAE monthly Accounting earnings

Continued on next page

 Acronym Long Description Author(s) Year, Journal Frequency Cat.Data Cat.Economic

FR Pension Funding Status Franzoni and Marin 2006, JF monthly Accounting composite accounting
BetaFP Frazzini-Pedersen Beta Frazzini and Pedersen 2014, JFE monthly Price other
High52 52 week high George and Hwang 2004, JF monthly Price momentum
IndMom Industry Momentum Grinblatt and Moskowitz 1999, JFE monthly Price momentum
PctAcc Percent Operating Accruals Hafzalla, Lundholm, Van Winkle 2011, AR yearly Accounting accruals
PctTotAcc Percent Total Accruals Hafzalla, Lundholm, Van Winkle 2011, AR yearly Accounting accruals
tang Tangibility Hahn and Lee 2009, JF yearly Accounting asset composition
Coskewness Coskewness Harvey and Siddique 2000, JF monthly Price risk
RoE net income / book equity Haugen and Baker 1996, JFE yearly Accounting profitability
VarCF Cash-flow to price variance Haugen and Baker 1996, JFE monthly Accounting cash flow risk
VolMkt Volume to market equity Haugen and Baker 1996, JFE monthly Trading volume
VolumeTrend Volume Trend Haugen and Baker 1996, JFE monthly Trading volume
58

AnalystRevision EPS forecast revision Hawkins, Chamberlin, Daniel 1984, FAJ monthly Analyst earnings
Mom12mOffSeason Momentum without the seasonal part Heston and Sadka 2008, JFE monthly Price momentum
MomOffSeason Off season long-term reversal Heston and Sadka 2008, JFE monthly Price momentum
MomOffSeason06YrPlus Off season reversal years 6 to 10 Heston and Sadka 2008, JFE monthly Price momentum
MomOffSeason11YrPlus Off season reversal years 11 to 15 Heston and Sadka 2008, JFE monthly Price momentum
MomOffSeason16YrPlus Off season reversal years 16 to 20 Heston and Sadka 2008, JFE monthly Price momentum
MomSeason Return seasonality years 2 to 5 Heston and Sadka 2008, JFE monthly Price momentum
MomSeason06YrPlus Return seasonality years 6 to 10 Heston and Sadka 2008, JFE monthly Price momentum
MomSeason11YrPlus Return seasonality years 11 to 15 Heston and Sadka 2008, JFE monthly Price momentum
MomSeason16YrPlus Return seasonality years 16 to 20 Heston and Sadka 2008, JFE monthly Price momentum
MomSeasonShort Return seasonality last year Heston and Sadka 2008, JFE monthly Price momentum
NOA Net Operating Assets Hirshleifer et al. 2004, JAE yearly Accounting asset composition

Continued on next page

 Acronym Long Description Author(s) Year, Journal Frequency Cat.Data Cat.Economic

dNoa change in net operating assets Hirshleifer, Hou, Teoh, Zhang 2004, JAE yearly Accounting investment
EarnSupBig Earnings surprise of big firms Hou 2007, RFS quarterly Accounting delayed processing
IndRetBig Industry return of big firms Hou 2007, RFS monthly Price delayed processing
PriceDelayRsq Price delay r square Hou and Moskowitz 2005, RFS monthly Price delayed processing
PriceDelaySlope Price delay coeff Hou and Moskowitz 2005, RFS monthly Price delayed processing
PriceDelayTstat Price delay SE adjusted Hou and Moskowitz 2005, RFS monthly Price delayed processing
STreversal Short term reversal Jegadeesh 1989, JF monthly Price short-term reversal
RevenueSurprise Revenue Surprise Jegadeesh and Livnat 2006, JFE quarterly Accounting sales growth
Mom12m Momentum (12 month) Jegadeesh and Titman 1993, JF monthly Price momentum
Mom6m Momentum (6 month) Jegadeesh and Titman 1993, JF monthly Price momentum
ChangeInRecommendation Change in recommendation Jegadeesh et al. 2004, JF monthly Analyst recommendation
OptionVolume1 Option to stock volume Johnson and So 2012, JFE monthly Trading volume
59

OptionVolume2 Option volume to average Johnson and So 2012, JFE monthly Trading volume
BetaTailRisk Tail risk beta Kelly and Jiang 2014, RFS monthly Price risk
fgr5yrLag Long-term EPS forecast La Porta 1996, JF monthly Analyst earnings
CF Cash flow to market Lakonishok, Shleifer, Vishny 1994, JF monthly Accounting valuation
MeanRankRevGrowth Revenue Growth Rank Lakonishok, Shleifer, Vishny 1994, JF yearly Accounting sales growth
RDS Real dirty surplus Landsman et al. 2011, AR yearly Accounting composite accounting
Tax Taxable income to income Lev and Nissim 2004, AR yearly Accounting tax
RDcap R&D capital-to-assets Li 2011, RFS yearly Accounting asset composition
zerotrade Days with zero trades Liu 2006, JFE monthly Trading liquidity
zerotradeAlt1 Days with zero trades Liu 2006, JFE monthly Trading liquidity
zerotradeAlt12 Days with zero trades Liu 2006, JFE monthly Trading liquidity
ChEQ Growth in book equity Lockwood and Prombutr 2010, JFR yearly Accounting investment

Continued on next page

 Acronym Long Description Author(s) Year, Journal Frequency Cat.Data Cat.Economic

EarningsStreak Earnings surprise streak Loh and Warachka 2012, MS monthly Accounting earnings
NumEarnIncrease Earnings streak length Loh and Warachka 2012, MS quarterly Accounting earnings
GrAdExp Growth in advertising expenses Lou 2014, RFS yearly Accounting investment
EntMult Enterprise Multiple Loughran and Wellman 2011, JFQA monthly Accounting valuation
CompositeDebtIssuance Composite debt issuance Lyandres, Sun and Zhang 2008, RFS yearly Accounting external financing
InvestPPEInv change in ppe and inv/assets Lyandres, Sun and Zhang 2008, RFS yearly Accounting investment
Frontier Efficient frontier index Nguyen and Swanson 2009, JFQA yearly Accounting valuation
GP gross profits / total assets Novy-Marx 2013, JFE yearly Accounting profitability
IntMom Intermediate Momentum Novy-Marx 2012, JFE monthly Price momentum
OPLeverage Operating leverage Novy-Marx 2010, ROF yearly Accounting other
Cash Cash to assets Palazzo 2012, JFE quarterly Accounting asset composition
BetaLiquidityPS Pastor-Stambaugh liquidity beta Pastor and Stambaugh 2003, JPE monthly Price liquidity
60

BPEBM Leverage component of BM Penman, Richardson and Tuna 2007, JAR monthly Accounting leverage
EBM Enterprise component of BM Penman, Richardson and Tuna 2007, JAR monthly Accounting valuation
NetDebtPrice Net debt to price Penman, Richardson and Tuna 2007, JAR monthly Accounting leverage
PS Piotroski F-score Piotroski 2000, AR yearly Accounting composite accounting
ShareIss1Y Share issuance (1 year) Pontiff and Woodgate 2008, JF monthly Accounting external financing
DelDRC Deferred Revenue Prakash and Sinha 2012, CAR yearly Accounting investment
OrderBacklog Order backlog Rajgopal, Shevlin, Venkatachalam 2003, RAS yearly Accounting sales growth
DelCOA Change in current operating assets Richardson et al. 2005, JAE yearly Accounting investment
DelCOL Change in current operating liabilities Richardson et al. 2005, JAE yearly Accounting external financing
DelEqu Change in equity to assets Richardson et al. 2005, JAE yearly Accounting investment
DelFINL Change in financial liabilities Richardson et al. 2005, JAE yearly Accounting external financing
DelLTI Change in long-term investment Richardson et al. 2005, JAE yearly Accounting investment

Continued on next page

 Acronym Long Description Author(s) Year, Journal Frequency Cat.Data Cat.Economic

DelNetFin Change in net financial assets Richardson et al. 2005, JAE yearly Accounting investment
TotalAccruals Total accruals Richardson et al. 2005, JAE yearly Accounting investment
BM Book to market using most recent ME Rosenberg, Reid, and Lanstein 1985, JF monthly Accounting valuation
Accruals Accruals Sloan 1996, AR yearly Accounting accruals
ChAssetTurnover Change in Asset Turnover Soliman 2008, AR yearly Accounting sales growth
ChNNCOA Change in Net Noncurrent Op Assets Soliman 2008, AR yearly Accounting investment
ChNWC Change in Net Working Capital Soliman 2008, AR yearly Accounting investment
ChInv Inventory Growth Thomas and Zhang 2002, RAS yearly Accounting investment
ChTax Change in Taxes Thomas and Zhang 2011, JAR quarterly Accounting tax
Investment Investment to revenue Titman, Wei and Xie 2004, JFQA yearly Accounting investment
realestate Real estate holdings Tuzel 2010, RFS yearly Accounting asset composition
AbnormalAccruals Abnormal Accruals Xie 2001, AR yearly Accounting accruals
61

FirmAgeMom Firm Age - Momentum Zhang 2004, JF monthly Price momentum

Table D.1: The table shows all available characteristics used, the author(s), the year and the journal of publication. In addition, this table shows the
update frequency, the data category as well as the economic category.
Table D.2: Hyperparameters

PPP DPPP
L1 penalty λ ∈ {0, 10−5 , 10−3 } λ ∈ {0, 10−5 , 10−3 }
Learning Rate 0.001 0.001
Dropout 0 D ∈ {0, 0.2, 0.4}
Batch Size 12 12
Epochs 200 200
Patience 20 20
Ensemble 0 5
Leaky ReLU − 0.01

This table gives the hyperparameters that we tune. The first column shows the hyperparameters for the
linear parametric portfolio policy (PPP). The second column shows the hyperparameters for the deep
parametric portfolio policy (DPPP).

62
Table D.3: Deep portfolio policy with different number of hidden layers

Layer 2 Layer 3 Layer 4 Layer 5

Utility 0.0510 0.0559 0.0548 0.0567
p-value(ULi − UL3 ) 0.0797 0.3460 0.3996

∑ |wi |/Nt ∗ 100 1.2532 1.1636 1.1648 0.9059

max wi ∗ 100 2.3479 2.1713 2.3620 2.2910
min wi ∗ 100 -2.2734 -2.1431 -2.3701 -2.3172
∑ wi I ( wi < 0) -8.5341 -7.8884 -7.8970 -6.0309
∑ I (wi < 0)/Nt 0.4838 0.4751 0.4704 0.4607
+
∑ |wi,t − wi,t −1 | 15.4961 14.1319 14.5830 11.8094

Mean 0.1111 0.1042 0.1159 0.1037

StdDev 0.1553 0.1392 0.1566 0.1374
Skew 0.2862 0.3584 0.6250 0.6940
Kurt 1.6373 1.0855 1.9909 1.4173
SR 2.4785 2.5934 2.5634 2.6152
p-value(SR Li − SR L3 ) 0.0824 0.3359 0.4053

FF5 + Mom α 0.0933 0.0863 0.0988 0.0885

StdErr (α) 0.0085 0.0076 0.0088 0.0077

This table shows out-of-sample estimates of the deep portfolio policies optimized for a mean-variance
investor with absolute risk aversion of five conditional on 157 firm characteristics. The deep models are
feed-forward neural networks with two (32, 16), three (32, 16, 8), four (32, 16, 8, 4) and five (32, 16, 8, 4, 2)
hidden layers (nodes), respectively. We use data from the Open Source Asset Pricing Dataset from January
1971 to December 2020. The columns labeled "Layer 2", "Layer 3", "Layer 4" and "Layer 5" show the statistics
of the deep parametric portfolio policy with two, three, four and five hidden layers, respectively. The first
rows show the utility of the investor as well as the bootstrapped one-sided p-value for the difference in
utility between the model with 3 layers and the other models. The second set of rows shows statistics on
portfolio weights averaged over time. These statistics include the average absolute portfolio weight, the
average maximum and minimum portfolio weights, the average sum of negative weights in the portfolio,
the average proportion of negative weights in the portfolio, and the turnover in the portfolio. The third set
of rows shows the first four moments of the final portfolio return distributions as well as the annualized
Sharpe ratios and the bootstrapped one-sided p-value for the difference in Sharpe ratios between the model
with 3 layers and the other models. The bottom panel shows the alphas and their standard errors with
respect to the Fama-French five-factor model extended to include the momentum factor.

63
Table D.4: Deep portfolio policy with different network architectures

PPP DPPP Frequency Category

Utility 0.0267 0.0469 0.0499 0.0473
p-value(UDPPP − Uarchitecture ) 0.0574 0.3910

∑ |wi |/Nt ∗ 100 0.5060 0.6057 0.6360 0.6355

max wi ∗ 100 2.0748 1.7260 1.7891 1.7235
min wi ∗ 100 -2.2097 -1.8370 -1.8926 -1.8203
∑ wi I ( wi < 0) -3.1475 -3.8665 -4.0847 -4.0813
∑ I (wi < 0)/Nt 0.4334 0.4411 0.4471 0.4478
+
∑ |wi,t − wi,t −1 | 3.9370 7.6984 8.3253 8.2927

Mean 0.0468 0.0701 0.0699 0.0670

StdDev 0.0897 0.0965 0.0898 0.0889
Skew -0.1451 1.0537 0.2517 0.2430
Kurt 1.8391 6.5084 1.7880 2.2713
SR 1.8070 2.5170 2.6985 2.6101
p-value(SR DPPP − SR architecture ) 0.0318 0.1581

FF5 + Mom α 0.0323 0.0559 0.0564 0.0526

StdErr (α) 0.0040 0.0051 0.0048 0.0046

This table shows out-of-sample estimates of the deep and linear portfolio policies optimized for a mean-
variance investor with absolute risk aversion of five conditional on 157 firm characteristics. The regular
portfolio policy is a linear specification of Equation (4), while the deep model is a feed-forward neural
network with three hidden layers and 32, 16, and eight nodes, respectively. We use data from the Open
Source Asset Pricing Dataset from January 1971 to December 2020. The columns labeled "PPP", "DPPP",
"Frequency" and "Category" show the statistics of the linear portfolio policy, deep portfolio policy, deep
portfolio policy with variables grouped by frequency, and deep portfolio policy with variables grouped
by category, respectively. The last to columns refer to different network architectures where the variables
are only interacted with variables of their own group in the first hidden layer. The first rows show the
utility of the investor as well as the bootstrapped one-sided p-value for the difference in utility between the
DPPP and the "Frequency" and "Category" models, respectively. The second set of rows shows statistics on
portfolio weights averaged over time. These statistics include the average absolute portfolio weight, the
average maximum and minimum portfolio weights, the average sum of negative weights in the portfolio,
the average proportion of negative weights in the portfolio, and the turnover in the portfolio. The third set
of rows shows the first four moments of the final portfolio return distributions as well as the annualized
Sharpe ratios and the bootstrapped one-sided p-value for the difference in Sharpe ratios between the DPPP
and the "Frequency" and "Category" models, respectively. The bottom panel shows the alphas and their
standard errors with respect to the Fama-French five-factor model extended to include the momentum
factor.

64
 Centre for Financial Research
CFR Working Paper Series Cologne

CFR Working Papers are available for download from www.cfr-cologne.de.

2023
No. Author(s) Title
23-01 F. Simon, S. Weibels, Deep Parametric Portfolio Policies
T. Zimmermann

2022
No. Author(s) Title
22-12 V. Agarwal, A. Cochardt, Birth Order and Fund Manager’s Trading Behavior: Role of
V. Orlov Sibling Rivalry

22-11 G. Cici, S. Gibson, N. Qin, The Performance of Corporate Bond Mutual Funds and the
A. Zhang Allocation of Underpriced New Issues

22-10 E. Theissen, C. Westheide One for the Money, Two for the Show? The Number of
Designated Market Makers and Liquidity

22-09 R. Campbell, P. Limbach, Once Bitten, Twice Shy: Failed Deals and Subsequent M&A
J. Reusche Cautiousness

22-08 M. Gehde-Trapp, L. The Effect of Sentiment on Institutional Investors: A Gender

Klingler Analysis

22-07 T. Bauckloh, V. Beyer, C. Does it Pay to Invest in Dirty Industries? – New Insights on the
Klein Shunned-Stock Hypothesis

22-06 J. Balthrop and G. Cici Conflicting Incentives in the Management of 529 Plans

22-05 I. T. Ivanov, T. Limits of Disclosure Regulation in the Municipal Bond Market

Zimmermann, N. W.
Heinrich

22-04 M. Ammann, A. Cochardt, Back to the Roots: Ancestral Origin and Mutual Fund Manager
S. Straumann, F. Weigert Portfolio Choice

22-03 A. Betzer, J. Gider, P. Do Financial Advisors Matter for M&A Pre-Announcement

Limbach Returns?

22-02 S. Lesmeister, P. Limbach, Indexing and the Performance-Flow Relation of Actively

P.R. Rau, F. Sonnenburg Managed Mutual Funds

22-01 T. Bauckloh, C. Klein, T. Under Pressure: The Link between Mandatory Climate
Pioch, F. Schiemann Reporting and Firms’ Carbon Performance
2021
No. Author(s) Title
21-11 V. Agarwal, H. Ren, K. Redemption in Kind and Mutual Fund Liquidity Management
Shen, H. Zhao

21-10 N.C. Brown, W. B. Elliott, News or noise: Mobile internet technology and stock market
R. Wermers, R. M. White activity

21-09 V. Agarwal, B. Barber, S. Private Company Valuations by Mutual Funds

Cheng, A. Hameed, A.
Yasuda

21-08 T.G. Bali, H. Beckmeyer, Option Return Predictability with Machine Learning and Big
M. Moerke, F. Weigert Data

21-07 F. Chabi-Yo, M. Multivariate Crash Risk

Huggenberger, F. Weigert

21-06 V. Agarwal, H. Aslan, L. Political Uncertainty and Household Stock Market Participation
Huang, H. Ren

21-05 G. Cici, P. Zhang On the Valuation Skills of Corporate Bond Mutual Funds

21-04 V. Agarwal, P. Hanouna, Do ETFs Increase the Commonality in Liquidity of Underlying

R. Moussawi, C.W. Stahel Stocks?

21-03 M. Saglam, T. Tuzun, R. Do ETFs Increase Liquidity?

Wermers

21-02 C. Andres, D. Bazhutov, Does Speculative News Hurt Productivity? Evidence from
D. Cumming, P. Limbach Takeover Rumors

21-01 T.G. Bali, F. Weigert Hedge Funds and the Positive Idiosyncratic Volatility Effect

This document only covers the most recent CFR Working Papers. A full list can be
found at www.cfr-cologne.de.
Cfr/University of cologne
Albertus-Magnus-Platz
D-50923 Cologne
Fon +49(0)221-470-6995
Fax +49(0)221-470-3992
Kempf@cfr-Cologne.de
www.cfr-cologne.de