How does model adoption affect its performance in stock price prediction

Abstract

This thesis examines the relationship between adoption of machine learning models and their performance in stock return prediction. The real-time performance of four machine learning models — neural networks, long short-term memory networks, temporal fusion transformers and support vector classifiers — are evaluated from 2010 to 2024 with daily data from the U.S. stock market. The adoption rate is proxied by the number of publications in academic journals that are concerned with using these models for predicting stock returns. The results show no statistically significant relationship between adoption and performance for any of the models, even when controlling for the changes in market efficiency approximated by the Hurst exponent. All models achieve better directional accuracy than random guessing, but none consistently outperform the buy and hold returns from S&P 500. Support vector classifiers show the best results in terms of returns and directional accuracy.