Uncovering Nonlinear Patterns in Stock Returns Using Gated Recurrent Unit (GRU) Model

Abstract

The escalating complexity of financial markets has driven the adoption of advanced machine learning techniques, such as Gated Recurrent Unit (GRU) models, to uncover nonlinear patterns in stock returns that traditional linear methods fail to capture. While the EMH posits that stock prices follow a random walk, empirical evidence increasingly challenges this assumption, revealing nonlinear dependencies. The GRU achieves an R² of 0.0131, explaining only 0.13% of the variance in log returns, which underscores the inherent difficulty of forecasting noisy financial data. The observed pattern suggests marginal but exploitable predictability in returns, valuable in high-frequency trading. GRU models can uncover subtle nonlinear dependencies, enhancing both risk management and dynamic asset allocation. This highlights their practical utility in modern quantitative finance for capturing small but impactful market inefficiencies.