Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 94 of the 94 references it cites were found in a registry or an index; 6 of the 42 sources its claims cite were checked for support.

Empirical asset pricing via machine learning: evidence from the European stock market

Open access
Drobetz, Wolfgang, Otto, Tizian · 2021 · Journal of Asset Management

Abstract

This paper evaluates the predictive performance of machine learning methods in forecasting European stock returns. Compared to a linear benchmark model, interactions and nonlinear effects help improve the predictive performance. But machine learning models must be adequately trained and tuned to overcome the high dimensionality problem and to avoid overfitting. Across all machine learning methods, the most important predictors are based on price trends and fundamental signals from valuation ratios. However, the models exhibit substantial variation in statistical predictive performance that translate into pronounced differences in economic profitability. The return and risk measures of long-only trading strategies indicate that machine learning models produce sizeable gains relative to our benchmark. Neural networks perform best, also after accounting for transaction costs. A classification-based portfolio formation, utilizing a support vector machine that avoids estimating stock-level expected returns, performs even better than the neural network architecture.

Overview

CoChat Summary

TL;DR

This paper evaluates machine learning methods for forecasting European stock returns against a linear Fama–MacBeth benchmark using 22 firm-level predictors, and finds that interactions and nonlinear effects improve predictive performance when models are adequately trained and tuned. Neural networks perform best among expected-return-based approaches, but a classification-based support vector machine that avoids estimating stock-level expected returns performs even better.

Key findings

  • Machine learning models beat the linear benchmark on out-of-sample predictive R² (full sample): ols_full −0.52%, ols_pars 0.57%, elanet 1.21%, pcr 1.24%, pls 1.08%, rf 1.18%, gbrt 1.14%, nn_1 1.23%. Predictive slopes were closest to 1 for rf (1.03), gbrt (1.07), and nn_1 (0.98), versus 0.65 for ols_pars and 0.45 for ols_full.
  • The most important predictors across models are based on recent price trends (e.g., short-term reversal, momentum) and fundamental valuation signals (e.g., earnings-to-price, book-to-market); tree-based models concentrate weight on few variables, while dimension-reducing models spread importance more widely.
  • Long-only equal-weighted portfolios: nn_1 produced a monthly H–L spread of 3.26% (return ratio 1.04, annualized Sharpe 3.89, t-stat 13.96); the ols_pars portfolio reached a terminal value of €70.29 versus €107.18 for nn_1 before transaction costs, and €17.02 versus €26.01 after applying a conservative 57 bps monthly cost.
  • The SVM classification approach slightly outperformed nn_1: accuracy 12.63% vs. 12.10% (random classifier 10.04%), and higher terminal values (€113.57 vs. €107.18 equal-weighted before costs; €27.57 vs. €26.01 after costs).

Methods

  • Sample: all firms publicly listed in one of the nineteen Eurozone countries as of December 2020, monthly data from Thomson Reuters Datastream, January 1990–December 2020; average universe of 832 firms (range 56 to 1,012); average monthly excess return 0.51% (SD 4.55%); minimum €25 mill. market cap and at least 50 qualifying firms per month.
  • Predictors: 22 firm-level characteristics from Drobetz et al. (2019), plus 44 country/SIC dummies (parsimonious set = 66 covariates) and, in the full set, two-way interactions and second/third-order polynomials (341 covariates); characteristics winsorized at 1%/99% and rank-mapped to (−1, +1).
  • Models: OLS (parsimonious and full), elastic net, PCR, PLS, random forests, gradient boosted regression trees, feedforward neural networks (nn_1–nn_5), and a support vector machine for classification-based portfolio formation; rolling-window time-series cross-validation with 8 years training, 2 years validation, 1 year testing, annual re-estimation; evaluation via predictive slopes, predictive R², modified Diebold–Mariano tests, decile portfolio sorts, conditional CAPM/FF5M alphas, and long-only portfolio performance net of transaction costs.

Limitations

  • The paper states that neural networks are computationally intensive and can be specified in innumerable architectures, so the authors pre-specify five models and treat nn_1–nn_5 as "a conservative lower bound" for neural network performance; they also note that deeper networks deteriorated quickly, possibly due to overfitting on a relatively small dataset.
  • The paper notes that the DM test's weak error dependence assumption may be questionable due to potentially strong cross-sectional return dependence, and that pre-estimation variable selection based on full-sample importance metrics could introduce foresight bias.
  • Evident from the text: the applied 57 bps monthly transaction cost is described as an overly conservative discount that pushes most value-weighted portfolios below the benchmark, and the authors caution that statistical and economic performance may be only weakly associated (citing Leitch and Tanner 1991).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1057/s41260-021-00237-x

Cites

94 of 94 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Between-models comparison of model complexity over time. This figure presents the model complexity of selected machine learning models (elanet, pcr, pls, rf, and gbrt) introduced in the 'Forecast models' section, at each re-estimation date. As the models pertain to different families, the model complexity measure varies: the number of nonzero regression coefficients (for elanet), the optimal number of components included in the predictive regression (for pcr and pls), and the optimal depth (for rf) or the number of unique split variables (for gbrt) of the regression trees. The sample includes all firms that
Fig. 1 Between-models comparison of model complexity over time. This figure presents the model complexity of selected machine learning models (elanet, pcr, pls, rf, and gbrt) introduced in the 'Forecast models' section, at each re-estimation date. As the models pertain to different families, the model complexity measure varies: the number of nonzero regression coefficients (for elanet), the optimal number of components included in the predictive regression (for pcr and pls), and the optimal depth (for rf) or the number of unique split variables (for gbrt) of the regression trees. The sample includes all firms thatFrom the PDF
Image 2
From the PDF
Fig. 3 Between-models comparison of relative variable importance. This figure presents a heat map of average relative variable importance ranks of each machine learning model introduced in the 'Forecast models' section, which are obtained by ranking the relative contribution of each variable within a specific model, and summing the ranks across all models to obtain an overall rank (higher variable importance = higher rank). The rows are sorted in descending order based on overall rank. Darker cell colors denote greater importance for the respective variable to a model. The relative variable importance metrics are calculated based on a two-step approach: First, the absolute variable importance is computed as the reduction in ­ R 2 from
Fig. 3 Between-models comparison of relative variable importance. This figure presents a heat map of average relative variable importance ranks of each machine learning model introduced in the 'Forecast models' section, which are obtained by ranking the relative contribution of each variable within a specific model, and summing the ranks across all models to obtain an overall rank (higher variable importance = higher rank). The rows are sorted in descending order based on overall rank. Darker cell colors denote greater importance for the respective variable to a model. The relative variable importance metrics are calculated based on a two-step approach: First, the absolute variable importance is computed as the reduction in ­ R 2 fromFrom the PDF
Fig. 4 Relative variable importance over time. This figure shows the relative variable importance ranks of an exemplary selected machine learning model (ols_pars) introduced in the 'Forecast models' section, at each re-estimation date, which are obtained by ranking the relative contribution of each variable within a specific model (higher variable importance = higher rank). The relative variable importance metrics are calculated based on a two-step approach: First, the absolute variable importance is computed as the reduction in ­ R 2 from setting all values of a given predictor to zero within the training sample. Second, the absolute variable importance measures are normalized to
Fig. 4 Relative variable importance over time. This figure shows the relative variable importance ranks of an exemplary selected machine learning model (ols_pars) introduced in the 'Forecast models' section, at each re-estimation date, which are obtained by ranking the relative contribution of each variable within a specific model (higher variable importance = higher rank). The relative variable importance metrics are calculated based on a two-step approach: First, the absolute variable importance is computed as the reduction in ­ R 2 from setting all values of a given predictor to zero within the training sample. Second, the absolute variable importance measures are normalized toFrom the PDF
Panel A: Equal-weighted cumulative performance
Panel A: Equal-weighted cumulative performanceFrom the PDF
Fig. 5 Cumulative performance of expected return-sorted portfolios with monthly restructuring. This figure depicts the performance (logarithmic scale) of the market portfolio and each machine learning portfolio introduced in the 'Forecast models' section. Portfolio values are scaled to €1 at the beginning of January 2000 and presented for long- and short-only portfolios, for both the equal- and value-weighting scheme, and gross of transaction costs. The calculation of transaction costs and the choice of applied transaction costs are explained in the 'Investment portfolio performance' section.
Fig. 5 Cumulative performance of expected return-sorted portfolios with monthly restructuring. This figure depicts the performance (logarithmic scale) of the market portfolio and each machine learning portfolio introduced in the 'Forecast models' section. Portfolio values are scaled to €1 at the beginning of January 2000 and presented for long- and short-only portfolios, for both the equal- and value-weighting scheme, and gross of transaction costs. The calculation of transaction costs and the choice of applied transaction costs are explained in the 'Investment portfolio performance' section.From the PDF
Panel A: Neural network architecture
Panel A: Neural network architectureFrom the PDF
Panel C: Univariate optimization (learning rate shrinkage)
Panel C: Univariate optimization (learning rate shrinkage)From the PDF
Panel D: MSFE in training and validation sample (early stopping)
Panel D: MSFE in training and validation sample (early stopping)From the PDF
Fig. 6 Stylized visualizations | neural networks. This figure depicts four stylized visualizations that help explain the structure, functioning, and regularization of neural networks
Fig. 6 Stylized visualizations | neural networks. This figure depicts four stylized visualizations that help explain the structure, functioning, and regularization of neural networksFrom the PDF

Verification

Claims the paper makes

5 Valid · 6 Not valid · 27 Not checked · 194 Own results and interpretations

99 more claims

Create a free account to see all 232 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 63
  • Probably not a substantial claim. 62
  • The quote only partly states this claim. 26

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 81 references

Research works printed but not linked: 2

  1. Chordia, T., A. Goyal, and A. Saretto. 2017. p-Hacking: Evidence from Two Million Trading Strategies. Swiss Finance Institute Research Paper No. 17–37: 1–53.
  2. Cochrane, J.H. 2005. Asset Pricing. Revised. Princeton (NJ), US: Princeton University Press.

Cited by

See all 61

Metadata

Authors
Drobetz, W., & Otto, T.
Affiliations
Universität Hamburg (DE)
Corresponding author
Otto, Tizian
Venue
Journal of Asset Management
Published
2021-11-12
Volume
22
Issue
7
Pages
507-538
ISSN
  • 1470-8272 (ISSN-L)
  • 1479-179X
OpenAlex
W3213046477
Cited by
61
References
81
Citations by year
2026: 18 · 2025: 7 · 2024: 10 · 2023: 15 · 2022: 8 · 2021: 3
FWCI
7.46
Type
journal article
Language
en
Open access
Licence
Locations
  • Journal of Asset Management (publishedVersion, cc-by)
  • SSRN Electronic Journal (acceptedVersion)
  • Econstor (Econstor) (submittedVersion, cc-by)
  • RePEc: Research Papers in Economics (submittedVersion)
Topics
  • Social Sciences › Economics, Econometrics and Finance › Finance › Financial Markets and Investment Strategies
  • Social Sciences › Decision Sciences › Management Science and Operations Research › Stock Market Forecasting Methods
  • Social Sciences › Decision Sciences › Management Science and Operations Research › Risk and Portfolio Optimization
Keywords
asset pricing, stock return prediction, machine learning, European stock markets, artificial neural networks, support vector machine, nonlinear effects, price trends, valuation ratios, overfitting, high dimensionality, transaction costs

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-28, under version 3 of the claim definition (prompt extract-v3).

A second model labelled each claim on 2026-09-28 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Empirical asset pricing via machine learning: evidence from the European stock market. https://app.cochat.ai/papers/10.1057/s41260-021-00237-x
The paper
Drobetz, W., & Otto, T. (2021). Empirical asset pricing via machine learning: evidence from the European stock market. Journal of Asset Management. https://doi.org/10.1057/s41260-021-00237-x

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.