Overfitting is when you tune a strategy so tightly to historical data that it memorizes the past's random noise instead of learning a real, repeatable pattern. It looks brilliant on history and falls apart live — because noise doesn't repeat.

Why it happens

The more knobs you turn — indicators, thresholds, filters — the more ways you find to make a curve fit the past perfectly. Add enough parameters and you can “predict” any history exactly. That's not skill; it's memorization (look-ahead bias).

How to spot it

Red flags: a suspiciously smooth equity curve, performance that collapses on data the strategy never saw, and rules with oddly specific numbers (“buy at RSI 31.7”). If it only works in-sample, it's overfit (backtest vs forward test).

The best backtest in the room is usually the biggest liar. Real edges are lumpy and imperfect; only overfit ones look flawless.

How to avoid it

Keep strategies simple (fewer parameters, fewer ways to fit noise), always test on out-of-sample data, use walk-forward testing, and prefer robust ideas that work across many conditions over fragile ones tuned to one (walk-forward optimization, survivorship bias). A slightly worse backtest that holds up out-of-sample beats a perfect one that doesn't.