Assistant Professor of Economics Javier Pereira recently presented a paper and chaired a panel on market stability at the 15th Western Economic Association International Conference. The meeting took place at Keio University in Tokyo.
Pereira presented research co-authored with Elias W. Leavenworth Professor of Economics Chris Georges. In “Market Stability with Machine Learning Agents,” the authors used an agent-based model of a financial market to illustrate how attention to forecast model selection by traders affects asset price volatility and financial market stability. Traders in this model are critical of their own forecasting models of asset returns and perform ongoing model selection to improve them.
Pereira and Georges found via simulation that the addition of model selection and other regularization methods to the traders’ learning algorithms reduces but does not eliminate overfitting and resulting excess volatility. They also found that the relative performance of the different regularization methods depends on whether traders entertain simpler or more complex forecast rules.
These results suggest that even a high degree of attention to overfitting by traders who are engaged in data mining may not entirely eliminate destabilizing speculation. The results are also consistent with recent empirical findings that suggest “pockets of predictability” in asset returns.