Generating Parameter Sensitivity Distributions to Mitigate Snooping Bias
May 16, 2014 - Big Ideas
Is there a practical way to mitigating data snooping bias while exploring optimal parameter values? In his February 2014 paper entitled “Know Your System! – Turning Data Mining from Bias to Benefit through System Parameter Permutation” (the National Association of Active Investment Managers’ 2014 Wagner Award winner), Dave Walton outlines System Parameter Permutation (SPP) as an alternative to traditional in-sample/out-of-sample cross-validation and other more complex approaches to compensating for data snooping bias. SPP generates distributions of performance metrics for rules-based investment strategies by systematically collecting outcomes across plausible ranges of rule parameters (see the figure below). These distributions capture typical, optimal and worst-case outcomes. He explains how to apply SPP to estimate both long-run and short-run strategy performance and to test statistical significance. He provides an example that compares conventional in-sample/out-of-sample testing and SPP as applied to an asset rotation strategy based on relative momentum. Using logical arguments and examples, he concludes that: Keep Reading