Predicting stock returns with machine learning: Global versus sector models

Recent studies highlight the superior performance of non-linear machine learning models, such as neural networks, over traditional linear models in predicting cross-sectional stock returns. These models are capable of capturing complex non-linear interactions between predictive signals and future returns. This thesis researches whether sector-specific neural networks can detect sector-related relationships to outperform a global neural network. It evaluates the predictive power of these models at the stock level and in portfolios based on return forecasts, constructing long-short portfolios from the networks' sorted predictions. A global neural network model trained on the full sample of stocks dominates neural networks trained on individual GICS sectors in predicting the cross-section of US stock returns. Sector-specific neural networks fail to gain an advantage by capturing complex sector-specific interactions. They underperform the global neural network especially in the early out-of-sample period. The smaller sample size for each GICS sector requires a trade-off between model complexity and robust model estimation. Pooling the data for the global model solves this problem and supports the predictive power of neural networks for stock returns.