A new study reveals that feature engineering plays a decisive role in the effectiveness of machine learning investment strategies. Researchers constructed real-time ML strategies based on a comprehensive set of fundamental signals, finding that while these strategies generated economically and statistically significant out-of-sample returns, they underperformed compared to results using manually curated signals from existing literature.
Notably, a simple strategy that recursively ranked signals based on their historical predictive performance demonstrated superior out-of-sample results. The same pattern emerged when using signals based on historical returns. These findings highlight the critical importance of feature engineering and broader inductive biases in enhancing the economic value of machine learning investment approaches.
Published in the Journal of Financial Economics (classified as an A-level journal in Wuhan University's ranking system), this research provides valuable insights for quantitative finance professionals and ML practitioners seeking to optimize investment strategies.
Full study: https://www.sciencedirect.com/science/article/pii/S0304405X25001461
