This thesis investigates whether machine learning methods can outperform traditional asset pricing approaches in predicting individual stock returns. The analysis is motivated by the growing empirical evidence that financial markets exhibit nonlinear relationships, high-dimensional predictor sets, and time-varying dynamics that standard linear models are not designed to capture. To address this question, a structured model comparison is conducted on a panel of 1,000 U.S. stocks over the period 1957–2021, following the framework of Gu, Kelly, and Xiu (2020). Ten models of increasing complexity are estimated and compared across a test period spanning 1987 to 2021: ordinary least squares, LASSO, Ridge regression, Elastic Net, a Random Forest, a Gradient Boosting Machine, three feed-forward neural network architectures of increasing depth (NN1, NN2, NN3), and a Long Short-Term Memory (LSTM) network. All models are trained on the 1957–1974 period, tuned on a validation set covering 1975–1986, and evaluated on the test set without any look-ahead. Performance is measured using the out-of-sample R_OOS^2 defined in Gu, Kelly, and Xiu (2020), which benchmarks model forecasts against the naive prediction of zero excess return for every stock. The analysis was first validated on a preliminary sample of 100 stocks, which confirmed the directional coherence of the results at reduced scale before the full analysis was conducted on 1,000 stocks. The qualitative ordering of model performance is broadly consistent with the original paper, while the quantitative differences are explained by the greater homogeneity of the stock universe used in this thesis relative to the full CRSP database. Overall, the findings support the view that machine learning adds value in asset pricing primarily through regularization and nonlinear interactions, and that the choice between model classes should be informed by the available data rather than by complexity alone.

Asset Pricing and Machine Learning

FABRIS, MATTIA
2025/2026

Abstract

This thesis investigates whether machine learning methods can outperform traditional asset pricing approaches in predicting individual stock returns. The analysis is motivated by the growing empirical evidence that financial markets exhibit nonlinear relationships, high-dimensional predictor sets, and time-varying dynamics that standard linear models are not designed to capture. To address this question, a structured model comparison is conducted on a panel of 1,000 U.S. stocks over the period 1957–2021, following the framework of Gu, Kelly, and Xiu (2020). Ten models of increasing complexity are estimated and compared across a test period spanning 1987 to 2021: ordinary least squares, LASSO, Ridge regression, Elastic Net, a Random Forest, a Gradient Boosting Machine, three feed-forward neural network architectures of increasing depth (NN1, NN2, NN3), and a Long Short-Term Memory (LSTM) network. All models are trained on the 1957–1974 period, tuned on a validation set covering 1975–1986, and evaluated on the test set without any look-ahead. Performance is measured using the out-of-sample R_OOS^2 defined in Gu, Kelly, and Xiu (2020), which benchmarks model forecasts against the naive prediction of zero excess return for every stock. The analysis was first validated on a preliminary sample of 100 stocks, which confirmed the directional coherence of the results at reduced scale before the full analysis was conducted on 1,000 stocks. The qualitative ordering of model performance is broadly consistent with the original paper, while the quantitative differences are explained by the greater homogeneity of the stock universe used in this thesis relative to the full CRSP database. Overall, the findings support the view that machine learning adds value in asset pricing primarily through regularization and nonlinear interactions, and that the choice between model classes should be informed by the available data rather than by complexity alone.
2025
File in questo prodotto:
File Dimensione Formato  
FABRIS_MATTIA_889576.pdf

accesso aperto

Dimensione 1.09 MB
Formato Adobe PDF
1.09 MB Adobe PDF Visualizza/Apri

I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14247/31308