07. Structural Changes

How does one split data into training, validation, and test sets so as to avoid bias induced by structural changes? It’s not always better to use the most recent time period as the test set, sometimes it’s better to have a random sample of years in the middle of your dataset. You want there to be nothing SPECIAL about the hold-out set. If the test set was the quant meltdown or financial crisis—those would be special validation sets. If you test on those time periods, you would be left with the unanswerable question: was it just bad luck? There is still some value in a strategy that would work every year except during a financial crisis.

Alphas tend to decay over time, so one can argue that using the past 3 or 4 years as a hold out set is a tough test set. Lots of things work less and less over time because knowledge spreads and new data are disseminated. Broader dissemination of data causes alpha decay. A strategy that performed well when tested on a hold-out set of the past few years would be slightly more impressive than one tested on a less recent time period.