Practical Applications of Advances in Machine Learning Approaches for Financial Decision Making and Time-Series Analysis: A Panel Discussion

In Advances in Machine Learning Approaches for Financial Decision Making and Time-Series Analysis: A Panel Discussion, from the Spring 2023 issue of The Journal of Financial Data Science, Nino Antulov-Fantulin of Aisot Technologies AG and Petter Kolm of New York University conclude that even with machine learning (ML) models, predicting the behavior of financial markets is difficult. The nonstationary behavior of markets means that models must react quickly to regime shifts, and sufficient data for recalibration may be unavailable. The authors offer five pieces of advice for investment professionals seeking to use ML models: 1) a model's predictive performance should not be the main focus for designing trading strategies; 2) discerning patterns and predicting the future is very difficult because financial markets are mostly efficient; 3) good knowledge of finance, microeconomics, statistics, ML, and artificial intelligence (AI) is necessary; 4) using only historical data for training a model is not sufficient because future trades based on the model may affect future prices; and 5) deploying an ML-based model at scale is challenging. The conclusions reflect the remarks of panelists on a March 28, 2023, panel titled "Advances of ML Approaches for Financial Decision Making & Time Series Analysis" at the 2022 Applied Machine Learning Days (AMLD) organized by the Swiss Federal Institute of Technology. The discussion revealed shared skepticism toward the use of machine learning for forecasting asset returns in data-scarce settings. Instead, panelists are more comfortable when using other applications and where sufficient data are available, such as at intraday timescales. Collectively, they successfully apply ML to a large, diverse set of applications, ranging from designing high-frequency return forecasting and order execution strategies to calibrating agent-based models.