Research

Three tells of an overfit backtest

Three patterns we look for before a backtest is allowed into the staging environment.

One: smooth equity curves with no drawdown structure

Real strategies have texture. A backtest whose equity line is suspiciously straight, with no drawdown longer than a week, is almost always over-parameterised.

Two: parameter sensitivity that is too gentle

If the result barely changes when we perturb every parameter by ten percent, the strategy has been fitted to a small basin. We expect a real edge to be sensitive in some directions and stable in others.

Three: a missing failure case

Every honest strategy has a regime where it loses money. If a backtest cannot point to its bad weather, the test window was not long enough or the signal is overfit. Either way, the staging gate stays closed.