nep-big New Economics Papers
on Big Data
Issue of 2021‒05‒03
twenty papers chosen by
Tom Coupé
University of Canterbury

  1. Using Satellite Imagery and Deep Learning to Evaluate the Impact of Anti-Poverty Programs By Luna Yue Huang; Solomon Hsiang; Marco Gonzalez-Navarro
  2. Optimal Stopping via Randomized Neural Networks By Calypso Herrera; Florian Krach; Pierre Ruyssen; Josef Teichmann
  3. Applying Convolutional Neural Networks for Stock Market Trends Identification By Ekaterina Zolotareva
  4. Algorithm is Experiment: Machine Learning, Market Design, and Policy Eligibility Rules By Yusuke Narita; Kohei Yata
  5. The role of Common Agricultural Policy (CAP) in enhancing and stabilising farm income: an analysis of income transfer efficiency and the Income Stabilisation Tool By Biagini Luigi; Simone Severini
  6. Prediction of Food Production Using Machine Learning Algorithms of Multilayer Perceptron and ANFIS By Saeed Nosratabadi; Sina Ardabili; Zoltan Lakner; Csaba Mako; Amir Mosavi
  7. Trends in Opioid Use among Social Security Disability Insurance Applicants By April Wu; Paul O'Leary; Denise Hoffman
  8. Stochastic Gradient Variational Bayes and Normalizing Flows for Estimating Macroeconomic Models By Ramis Khbaibullin; Sergei Seleznev
  9. A year of pandemic: levels, changes and validity of well-being data from Twitter. Evidence from ten countries By Sarracino, Francesco; Greyling, Talita; O'Connor , Kelsey; Peroni, Chiara; Rossouw, Stephanie
  10. CATE meets ML: Conditional average treatment effect and machine learning By Jacob, Daniel
  11. Computational Performance of Deep Reinforcement Learning to find Nash Equilibria By Christoph Graf; Viktor Zobernig; Johannes Schmidt; Claude Kl\"ockl
  12. Addressing Sample Selection Bias for Machine Learning Methods By Dylan Brewer; Alyssa Carlson
  13. Mirror, mirror on the wall: Machine predictions and self-fulfilling prophecies By Bauer, Kevin; Gill, Andrej
  14. Constructing long-short stock portfolio with a new listwise learn-to-rank algorithm By Xin Zhang; Lan Wu; Zhixue Chen
  15. Artificial intelligence companies, goods and services: A trademark-based analysis By Shohei Nakazato; Mariagrazia Squicciarini
  16. Artificial intelligence and industrial innovation: Evidence from firm-level data By Rammer, Christian; Fernández, Gastón P.; Czarnitzki, Dirk
  17. Optimal Targeting in Fundraising: A Machine-Learning Approach By Tobias Cagala; Ulrich Glogowsky; Johannes Rincke; Anthony Strittmatter
  18. Loss-Based Variational Bayes Prediction By David T. Frazier; Ruben Loaiza-Maya; Gael M. Martin; Bonsoo Koo
  19. Forecasting Energy Commodity Prices: A Large Global Dataset Sparse Approach By Davide Ferrari; Francesco Ravazzolo; Joaquin Vespignani
  20. Distinguish the indistinguishable: a Deep Reinforcement Learning approach for volatility targeting models By Eric Benhamou; David Saltiel; Serge Tabachnik; Sui Kai Wong; François Chareyron

  1. By: Luna Yue Huang; Solomon Hsiang; Marco Gonzalez-Navarro
    Abstract: The rigorous evaluation of anti-poverty programs is key to the fight against global poverty. Traditional evaluation approaches rely heavily on repeated in-person field surveys to measure changes in economic well-being and thus program effects. However, this is known to be costly, time-consuming, and often logistically challenging. Here we provide the first evidence that we can conduct such program evaluations based solely on high-resolution satellite imagery and deep learning methods. Our application estimates changes in household welfare in the context of a recent anti-poverty program in rural Kenya. The approach we use is based on a large literature documenting a reliable relationship between housing quality and household wealth. We infer changes in household wealth based on satellite-derived changes in housing quality and obtain consistent results with the traditional field-survey based approach. Our approach can be used to obtain inexpensive and timely insights on program effectiveness in international development programs.
    Date: 2021–04
  2. By: Calypso Herrera; Florian Krach; Pierre Ruyssen; Josef Teichmann
    Abstract: This paper presents new machine learning approaches to approximate the solution of optimal stopping problems. The key idea of these methods is to use neural networks, where the hidden layers are generated randomly and only the last layer is trained, in order to approximate the continuation value. Our approaches are applicable for high dimensional problems where the existing approaches become increasingly impractical. In addition, since our approaches can be optimized using a simple linear regression, they are very easy to implement and theoretical guarantees can be provided. In Markovian examples our randomized reinforcement learning approach and in non-Markovian examples our randomized recurrent neural network approach outperform the state-of-the-art and other relevant machine learning approaches.
    Date: 2021–04
  3. By: Ekaterina Zolotareva
    Abstract: In this paper we apply a specific type ANNs - convolutional neural networks (CNNs) - to the problem of finding start and endpoints of trends, which are the optimal points for entering and leaving the market. We aim to explore long-term trends, which last several months, not days. The key distinction of our model is that its labels are fully based on expert opinion data. Despite the various models based solely on stock price data, some market experts still argue that traders are able to see hidden opportunities. The labelling was done via the GUI interface, which means that the experts worked directly with images, not numerical data. This fact makes CNN the natural choice of algorithm. The proposed framework requires the sequential interaction of three CNN submodels, which identify the presence of a changepoint in a window, locate it and finally recognize the type of new tendency - upward, downward or flat. These submodels have certain pitfalls, therefore the calibration of their hyperparameters is the main direction of further research. The research addresses such issues as imbalanced datasets and contradicting labels, as well as the need for specific quality metrics to keep up with practical applicability. This paper is the full text of the research, presented at the 20th International Conference on Artificial Intelligence and Soft Computing Web System (ICAISC 2021)
    Date: 2021–04
  4. By: Yusuke Narita; Kohei Yata
    Abstract: Algorithms produce a growing portion of decisions and recommendations both in policy and business. Such algorithmic decisions are natural experiments (conditionally quasi-randomly assigned instruments) since the algorithms make decisions based only on observable input variables. We use this observation to develop a treatment-effect estimator for a class of stochastic and deterministic algorithms. Our estimator is shown to be consistent and asymptotically normal for well-defined causal effects. A key special case of our estimator is a high-dimensional regression discontinuity design. The proofs use tools from differential geometry and geometric measure theory, which may be of independent interest. The practical performance of our method is first demonstrated in a high-dimensional simulation resembling decision-making by machine learning algorithms. Our estimator has smaller mean squared errors compared to alternative estimators. We finally apply our estimator to evaluate the effect of Coronavirus Aid, Relief, and Economic Security (CARES) Act, where more than \$10 billion worth of relief funding is allocated to hospitals via an algorithmic rule. The estimates suggest that the relief funding has little effects on COVID-19-related hospital activity levels. Naive OLS and IV estimates exhibit substantial selection bias.
    Date: 2021–04
  5. By: Biagini Luigi (Tutor); Simone Severini (Tutor)
    Abstract: Since its inception, the E.U.'s Common Agricultural Policy (CAP) aimed at ensuring an adequate and stable farm income. While recognizing that the CAP pursues a larger set of objectives, this thesis focuses on the impact of the CAP on the level and the stability of farm income in Italian farms. It uses microdata from a high standardized dataset, the Farm Accountancy Data Network (FADN), that is available in all E.U. countries. This allows if perceived as useful, to replicate the analyses to other countries. The thesis first assesses the Income Transfer Efficiency (i.e., how much of the support translate to farm income) of several CAP measures. Secondly, it analyses the role of a specific and relatively new CAP measure (i.e., the Income Stabilisation Tool - IST) that is specifically aimed at stabilising farm income. The assessment of the potential use of Machine Learning procedures to develop an adequate ratemaking in IST. These are used to predict indemnity levels because this is an essential point for a similar insurance scheme. The assessment of ratemaking is challenging: indemnity distribution is zero-inflated, not-continuous, right-skewed, and several factors can potentially explain it. We address these problems by using Tweedie distributions and three Machine Learning procedures. The objective is to assess whether this improves the ratemaking by using the prospective application of the Income Stabilization Tool in Italy as a case study. We look at the econometric performance of the models and the impact of using their predictions in practice. Some of these procedures efficiently predict indemnities, using a limited number of regressors, and ensuring the scheme's financial stability.
    Date: 2021–04
  6. By: Saeed Nosratabadi; Sina Ardabili; Zoltan Lakner; Csaba Mako; Amir Mosavi
    Abstract: Advancing models for accurate estimation of food production is essential for policymaking and managing national plans of action for food security. This research proposes two machine learning models for the prediction of food production. The adaptive network-based fuzzy inference system (ANFIS) and multilayer perceptron (MLP) methods are used to advance the prediction models. In the present study, two variables of livestock production and agricultural production were considered as the source of food production. Three variables were used to evaluate livestock production, namely livestock yield, live animals, and animal slaughtered, and two variables were used to assess agricultural production, namely agricultural production yields and losses. Iran was selected as the case study of the current study. Therefore, time-series data related to livestock and agricultural productions in Iran from 1961 to 2017 have been collected from the FAOSTAT database. First, 70% of this data was used to train ANFIS and MLP, and the remaining 30% of the data was used to test the models. The results disclosed that the ANFIS model with Generalized bell-shaped (Gbell) built-in membership functions has the lowest error level in predicting food production. The findings of this study provide a suitable tool for policymakers who can use this model and predict the future of food production to provide a proper plan for the future of food security and food supply for the next generations.
    Date: 2021–04
  7. By: April Wu; Paul O'Leary; Denise Hoffman
    Abstract: Several factors suggest that opioid use may be common among applicants to Social Security Disability Insurance (SSDI): the prevalence of opioid use, the suspected link between opioid use and declining rates of work, and the large share of new SSDI awardees who have conditions associated with opioid use. However, research-ready data on opioid use by these applicants, who are generally not eligible for Medicare, have not been available. SSDI applicants are required to report their medications, but they often do so in an open-ended text field, which requires additional coding before analysis. Our study is the first to provide statistics on opioid use among SSDI applicants. We used an innovative machine-learning method to identify opioids in medication text fields in SSDI administrative data. Specifically, we examined the prevalence of reported opioid use in a 30 percent random sample of initial-level SSDI applications stored in the Social Security AdministrationÕs Structured Data Repository (SDR) from 2007 through 2017, considering differences by demographic and other factors. We supplemented the SDR with two SSA administrative data sources: the Disability Analysis File, which provided award information, and the Numerical Identification System, which provided information on deaths. Using these sources, we produced statistics on the association between (1) opioid use among SSDI applicants and (2) SSDI award and death. Understanding the prevalence of reported opioid use among these individuals and the association between opioid use and later SSDI application outcomes may help in forecasting the future composition of the SSDI caseload. The paper found that: Over the 11-year analysis period, more than 30 percent of SSDI applicants reported using one or more opioids. This is higher than the rate of opioid use in the general population (29 versus 19 percent in 2016).Reported rates of opioid use among SSDI applicants varied over the analysis period. Rates increased from 2007 to a peak of 32 percent in 2012, followed by a decline to the period low of 26 percent in 2017.Reported opioid use varied by age and demographic characteristics. SSDI applicants ages 40Ð49 were the most likely age group to report opioid use; women were 3-4 percentage points more likely to report opioid use than men; and people with some college were the most likely education group to report opioid use.Reported opioid use is also correlated with application type. SSDI-only applicants who reported opioid use were 4-6 percentage points more likely to report opioid use than concurrent SSDI and SSI applicants.Reported opioid use varied greatly between geographic areas. Applicants from Rhode Island, Massachusetts, and Washington, DC, reported lower-than-average rates of opioid use in 2007 and consistently throughout the analysis period. Conversely, applicants from Delaware, Nevada, and Michigan consistently reported the highest rates of opioid use.There was a positive and statistically significant association between (1) reported opioid use and SSDI awards and (2) reported opioid use SSDI award and death. These associations do not demonstrate a causal relationship. The policy implications of the findings are: Application for SSDI provides an opportunity to identify opioid users, understand more about the nature of their use, and, if warranted, connect them with helpful services and supports.Given the prevalence of reported opioid use among SSDI applicants, our study may open the door to future research on how opioid use affects post-adjudication well-beingÑfor example, applicantsÕ employment outcomes after they are awarded or denied SSDI. Future studies might also consider tracking the trajectory of opioid use from application through award using Medicare data.This study also serves as a template for using previously untapped information in SSA administrative files. The machine learning approach used to identify opioid use in free-form text could be applied to other key indicators of interest to SSA.
    Date: 2021–03
  8. By: Ramis Khbaibullin (Bank of Russia, Russian Federation); Sergei Seleznev (Bank of Russia, Russian Federation)
    Abstract: We illustrate the ability of the stochastic gradient variational Bayes algorithm, which is a very popular machine learning tool, to work with macrodata and macromodels. Choosing two approximations (mean-field and normalizing flows), we test properties of algorithms for a set of models and show that these models can be estimated fast despite the presence of estimated hyperparameters. Finally, we discuss the difficulties and possible directions of further research.
    Keywords: Stochastic gradient variational Bayes, normalizing flows, mean-field approximation, sparse Bayesian learning, BVAR, Bayesian neural network, DFM.
    JEL: C11 C32 C32 C45 E17
    Date: 2020–10
  9. By: Sarracino, Francesco; Greyling, Talita; O'Connor , Kelsey; Peroni, Chiara; Rossouw, Stephanie
    Abstract: In this article we describe how well-being changed during 2020 in ten countries, namely Australia, Belgium, France, Germany, Great Britain, Italy, Luxembourg, New Zealand, South Africa, and Spain. Our measure of well-being is the Gross National Happiness (GNH), a country-level index built applying sentiment analysis to data from Twitter. Our aim is to describe how GNH changed during the pandemic within countries, to assess its validity as a measure of well-being, and to analyse its correlates. We take advantage of a unique data-set made of daily observations about GNH, generalized trust and trust in national institutions, fear concerning the economy, loneliness, infection rate, policy stringency and distancing. To assess the validity of data sourced from Twitter, we exploit various sources of survey data, such as Eurobarometer and consumer satisfaction, and big data, such as Google Trends. Results indicate that sentiment analysis of Tweets an provide reliable and timely information on well-being. This can be particularly useful to timely inform decision-making.
    Keywords: happiness,Covid-19,Big Data,Twitter,Sentiment Analysis,well-being,public policy,trust,fear,loneliness
    JEL: C55 I10 I31 H12
    Date: 2021
  10. By: Jacob, Daniel
    Abstract: For treatment effects - one of the core issues in modern econometric analysis - prediction and estimation are flip-sides of the same coin. As it turns out, machine learning methods are the tool for generalized prediction models. Combined with econometric theory allows us to estimate not only the average but a personalized treatment effect - the conditional average treatment effect (CATE). In this tutorial, we give an overview of novel methods, explain them in detail, and apply them via Quantlets in real data applications. We study the effect that microcredit availability has on the amount of money borrowed and if the 401(k) pension plan eligibility has an impact on net financial assets, as two empirical examples. The presented toolbox of methods contains metalearners, like the Doubly-Robust, the R-, T- and X-learner, and methods that are specially designed to estimate the CATE like the causal BART and the generalized random forest. In both, the microcredit and the 401(k) example, we find a positive treatment effect for all observations but diverse evidence of treatment effect heterogeneity. An additional simulation study, where the true treatment effect is known, allows us to compare the different methods and to observe patterns and similarities.
    Keywords: Causal Inference,CATE,Machine Learning,Tutorial
    JEL: C00
    Date: 2021
  11. By: Christoph Graf; Viktor Zobernig; Johannes Schmidt; Claude Kl\"ockl
    Abstract: We test the performance of deep deterministic policy gradient (DDPG), a deep reinforcement learning algorithm, able to handle continuous state and action spaces, to learn Nash equilibria in a setting where firms compete in prices. These algorithms are typically considered model-free because they do not require transition probability functions (as in e.g., Markov games) or predefined functional forms. Despite being model-free, a large set of parameters are utilized in various steps of the algorithm. These are e.g., learning rates, memory buffers, state-space dimensioning, normalizations, or noise decay rates and the purpose of this work is to systematically test the effect of these parameter configurations on convergence to the analytically derived Bertrand equilibrium. We find parameter choices that can reach convergence rates of up to 99%. The reliable convergence may make the method a useful tool to study strategic behavior of firms even in more complex settings. Keywords: Bertrand Equilibrium, Competition in Uniform Price Auctions, Deep Deterministic Policy Gradient Algorithm, Parameter Sensitivity Analysis
    Date: 2021–04
  12. By: Dylan Brewer (School of Economics, Georgia Institute of Technology); Alyssa Carlson (Department of Economics, University of Missouri)
    Abstract: We study approaches for adjusting machine learning methods when the training sample differs from the prediction sample on unobserved dimensions. The machine learning literature predominately assumes selection only on observed dimensions. Common suggestions are to re-weight or control for variables that influence selection as solutions to selection on observables. Simulation results indicate that common machine learning practices such as re-weighting or controlling for variables that influence selection into the training or testing sample often worsens sample selection bias. We suggest two control-function approaches that remove the effects of selection bias before training and find that they reduce meansquared prediction error in simulations with a high degree of selection. We apply these approaches to predicting the vote share of the incumbent in gubernatorial elections using previously observed re-election bids. We find that ignoring selection on unobservables leads to substantially higher predicted vote shares for the incumbent than when the control function approach is used.
    Date: 2021
  13. By: Bauer, Kevin; Gill, Andrej
    Abstract: We show that disclosing machine predictions to affected parties can trigger self-fulfilling prophecies. In an investment game, we experimentally vary investors' and recipients' access to a machine prediction about recipients' likelihood to pay back an investment. Recipients who privately learn about an incorrect machine prediction alter their behavior in the direction of the prediction. Furthermore, when recipients learn that an investor has disregarded a machine prediction of no-repayment, this further lowers the repayment amount. We interpret these findings as evidence that transparency regarding machine predictions can alter recipients' beliefs about what kind of person they are and what investors expect of them. Our results indicate that providing increased access to machine predictions as an isolated measure to alleviate accountability concerns may have unintended negative consequences for organizations by possibly changing customer behavior.
    Keywords: algorithmic transparency,algorithmic decision support,human-machine interaction
    JEL: C91 D80 D91 O33
    Date: 2021
  14. By: Xin Zhang; Lan Wu; Zhixue Chen
    Abstract: Factor strategies have gained growing popularity in industry with the fast development of machine learning. Usually, multi-factors are fed to an algorithm for some cross-sectional return predictions, which are further used to construct a long-short portfolio. Instead of predicting the value of the stock return, emerging studies predict a ranked stock list using the mature learn-to-rank technology. In this study, we propose a new listwise learn-to-rank loss function which aims to emphasize both the top and the bottom of a rank list. Our loss function, motivated by the long-short strategy, is endogenously shift-invariant and can be viewed as a direct generalization of ListMLE. Under different transformation functions, our loss can lead to consistency with binary classification loss or permutation level 0-1 loss. A probabilistic explanation for our model is also given as a generalized Plackett-Luce model. Based on a dataset of 68 factors in China A-share market from 2006 to 2019, our empirical study has demonstrated the strength of our method which achieves an out-of-sample annual return of 38% with the Sharpe ratio being 2.
    Date: 2021–04
  15. By: Shohei Nakazato; Mariagrazia Squicciarini
    Abstract: This work proposes an experimental methodology to identify and measure artificial intelligence (AI)-related trademarks. It aims to shed light on the extent to which (new) companies and products appearing on the market rely on, exploit or propose AI-related goods and services, and to help identify the companies and organisations that are active in the AI space. The paper finds evidence that AI-related goods and services have expanded in consumer markets in recent years. Companies and other economic agents appear to register AI-related trademarks primarily to protect computer-related products and/or services, especially software, audio-visual devices and for analytical purposes. Important trademark activities related to AI also emerge in the education space, with AI-related keywords being frequently associated with educational services as well as classes, publications, workshops and online material.
    Keywords: Artificial Intelligence, goods and services, markets, trademarks
    JEL: O31 O34 L25
    Date: 2021–05–04
  16. By: Rammer, Christian; Fernández, Gastón P.; Czarnitzki, Dirk
    Abstract: Artificial Intelligence (AI) represents a set of techniques that enable new ways of innovation and allows firms to offer new features of products and services, to improve production, marketing and administration processes, and to introduce new business models. This paper analyses the extent to which the use of AI contributes to the innovation performance of firms. Based on firm-level data from the German part of the Community Innovation Survey (CIS) 2018, we examine the contribution of different AI methods and applications to product and process innovation outcomes. The representative nature of the survey allows extrapolating the findings to the macroeconomic level. The results show that 5.8% of firms in Germany were actively using AI in their business operations or products and services in 2019. The use of AI generated additional sales with world-first product innovations in these firms of about €16 billion, which corresponds to 18% of total sales of world-first innovations in the German business sector. Firms that developed AI by combining in-house and external resources obtained significantly higher innovation results. The same is true for firms that apply AI in a broad way and have already several years of experience in using AI.
    Keywords: Artificial Intelligence,Innovation,CIS data,Germany
    JEL: O14 O31 O32 O33 L25 M15
    Date: 2021
  17. By: Tobias Cagala; Ulrich Glogowsky; Johannes Rincke; Anthony Strittmatter
    Abstract: Ineffective fundraising lowers the resources charities can use for goods provision. We combine a field experiment and a causal machine-learning approach to increase a charity’s fundraising effectiveness. The approach optimally targets fundraising to individuals whose expected donations exceed solicitation costs. Among past donors, optimal targeting substantially increases donations (net of fundraising costs) relative to bench-marks that target everybody or no one. Instead, individuals who were previously asked but never donated should not be targeted. Further, the charity requires only publicly available geospatial information to realize the gains from targeting. We conclude that charities not engaging in optimal targeting waste resources.
    Keywords: fundraising, charitable giving, gift exchange, targeting, optimal policy learning, individualized treatment rules
    JEL: C93 D64 H41 L31 C21
    Date: 2021
  18. By: David T. Frazier; Ruben Loaiza-Maya; Gael M. Martin; Bonsoo Koo
    Abstract: We propose a new method for Bayesian prediction that caters for models with a large number of parameters and is robust to model misspecification. Given a class of high-dimensional (but parametric) predictive models, this new approach constructs a posterior predictive using a variational approximation to a loss-based, or Gibbs, posterior that is directly focused on predictive accuracy. The theoretical behavior of the new prediction approach is analyzed and a form of optimality demonstrated. Applications to both simulated and empirical data using high-dimensional Bayesian neural network and autoregressive mixture models demonstrate that the approach provides more accurate results than various alternatives, including misspecified likelihood-based predictions.
    Date: 2021–04
  19. By: Davide Ferrari (Free University of Bozen-Bolzano, Italy); Francesco Ravazzolo (Free University of Bozen-Bolzano, Italy; BI Norwegian Business School, Norway); Joaquin Vespignani (University of Tasmania, Tasmanian School of Business and Economics, Australia)
    Abstract: This paper focuses on forecasting quarterly nominal global energy prices of commodities, such as oil, gas and coal, using the Global VAR dataset proposed by Mohaddes and Raissi (2018). This dataset includes a number of potentially informative quarterly macroeconomic variables for the 33 largest economies, overall accounting for more than 80% of the global GDP. To deal with the information on this large database, we apply dynamic factor models based on a penalized maximum likelihood approach that allows to shrink parameters to zero and to estimate sparse factor loadings. The estimated latent factors show considerable sparsity and heterogeneity in the selected loadings across variables. When the model is extended to predict energy commodity prices up to four periods ahead, results indicate larger predictability relative to the benchmark random walk model for 1-quarter ahead for all energy commodities and up to 4 quarters ahead for gas prices. Our model also provides superior forecasts than machine learning techniques, such as elastic net, LASSO and random forest, applied to the same database.
    Keywords: Energy Prices; Forecasting; Dynamic Factor model; Sparse Estimation; Penalized Maximum Likelihood.
    JEL: C1 C5 C8 E3 Q4
    Date: 2021–04
  20. By: Eric Benhamou (LAMSADE - Laboratoire d'analyse et modélisation de systèmes pour l'aide à la décision - Université Paris Dauphine-PSL - PSL - Université Paris sciences et lettres - CNRS - Centre National de la Recherche Scientifique); David Saltiel; Serge Tabachnik; Sui Kai Wong; François Chareyron
    Abstract: Can an agent efficiently learn to distinguish extremely similar financial models in an environment dominated by noise and regime changes? Standard statistical methods based on averaging or ranking models fail precisely because of regime changes and noisy environments. Additional contextual information in Deep Reinforcement Learning (DRL), helps training an agent distinguish different financial models whose time series are very similar. Our contributions are four-fold: (i) we combine model-based and modelfree Reinforcement Learning (RL). The last model-free RL allows us selecting the different models, (ii) we present a concept, called "walk-forward analysis", which is defined by successive training and testing based on expanding periods, to assert the robustness of the resulting agent, (iii) we present a method based on the importance of features that looks like the one in gradient boosting methods and is based on features sensitivities, (iv) last but not least, we introduce the concept of statistical difference significance based on a two-tailed T-test, to highlight the ways in which our models differ from more traditional ones. Our experimental results show that our approach outperforms the benchmarks in almost all evaluation metrics commonly used in financial mathematics, namely net performance, Sharpe ratio, Sortino, maximum drawdown, maximum drawdown over volatility.
    Keywords: Features sensitivity,Walk forward,Portfolio allocation,Model-free,Model-based,Deep Reinforcement learning
    Date: 2021–04–21

This nep-big issue is ©2021 by Tom Coupé. It is provided as is without any express or implied warranty. It may be freely redistributed in whole or in part for any purpose. If distributed in part, please include this notice.
General information on the NEP project can be found at For comments please write to the director of NEP, Marco Novarese at <>. Put “NEP” in the subject, otherwise your mail may be rejected.
NEP’s infrastructure is sponsored by the School of Economics and Finance of Massey University in New Zealand.