Extreme Value Theory and Backtest Overfitting in Finance

In order to identify potentially profitable investment strategies, hedge funds and asset managers can use historical market data to simulate a strategy's performance, a process known as backtesting. While the abundance of historical stock price data and powerful computing technologies has made it feasible to run millions of simulations in a short period of time, this process may produce statistically insignificant results in the form of false positives. As the number of configurations of a strategy increases, it becomes more likely that some of the configurations will perform well by chance alone. The phenomenon of backtest overfitting occurs when a model interprets market idiosyncrasies as signal rather than noise, and is often not taken into account in the strategy selection process. As a result, the finance industry and academic literature are rife with skill-less strategies that have no capability of beating the market. This paper explores the development of a minimum criterion that managers and investors can use during the backtesting process in order to increase confidence that a strategy's performance is not the result of pure chance. To do this we will use extreme value theory to determine the probability of observing a specific result, or something more extreme than this result, given that multiple configurations of a strategy were tested.