Evolving RNNs for Stock Forecasting: A Low Parameter Efficient Alternative to Transformers

Zimeng Lyu, Devroop Kar, Matthew Simoni, Rohaan Nadeem, Avinash Bhojanapalli, Hao Zhang, Travis Desell

Abstract: Stock return forecasting is a critical application of time series forecasting in finance, facilitating informed trading and management decisions that can lead to substantial returns. However, for large investment portfolios, designing and fine-tuning models for individual stock predictions is time-consuming and computationally intensive. In this work, we propose using the neuroevolution-based neural architecture search algorithm, Evolutionary eXploration of Augmenting Memory Models (EXAMM), to evolve recurrent neural networks (RNNs) for stock return prediction. We compare the prediction performance of these evolved RNNs with that of state-of-the-art attention-based Transformer and deep learning models. Our results indicate that EXAMM-evolved RNNs outperform or achieve comparable performance with these models across 50 multivariate stock datasets and a combined high-dimensional dataset with 300 input features and 50 outputs. Additionally, they require orders of magnitude fewer parameters and can be evolved and operate efficiently using a minimal 8-core CPU configuration as opposed to expensive GPUs.