Return Forecasting and Trading Profitability Across Equity Portfolios: A Neural Network Approach Versus Traditional Benchmarks

Abstract

While research increasingly suggests that stock returns may be predictable, particularly through nonlinear methods such as machine learning, the evidence on their practical advantage remains mixed. Critics argue that these models do not consistently outperform traditional statistical approaches, and their profitability is often uncertain. This thesis contributes to the debate by evaluating whether an Artificial Neural Network (ANN) can outperform a linear Autoregressive (AR) model and a passive Buy and Hold (BH) strategy across four U.S. equity portfolios: Small Cap, Large Cap, Value, and Growth. Using four assets, two prediction targets, seventeen input configurations, and two training and testing samples, the study assesses 272 different ANN and AR model configurations. Predictive accuracy is measured using the Henriksson-Merton test, and economic performance is evaluated through the Sortino ratio and CAPM alpha. All performance measures are subjected to formal statistical testing. The results suggest that while ANN models show potential in specific contexts, they are not universally superior. Their effectiveness depends on the asset class, model design, prediction target, and the performance measure used.