Market stability with machine learning agents

We consider the effect of adaptive model selection and regularization by agents on price volatility and market stability in a simple agent-based model of a financial market. The agents base their trading behavior on forecasts of future returns, which they update adaptively and asynchronously through a process of model selection, estimation, and prediction. The addition of model selection and regularization methods to the traders’ learning algorithm is shown to reduce but not eliminate overfitting and resulting excess volatility. Our results suggest that even a high degree of attention to overfitting on the part of traders who are engaged in data mining is unlikely to entirely eliminate destabilizing speculation. They also accord well with the empirical “sparse signals” and “pockets of predictability” findings of Chinco et al. (2019) and Farmer et al. (2019).