Why do backtests fail in live trading?

The usual killer is curve-fitting: the strategy's parameters were tuned — deliberately or by trial and error — until they fit the historical data's noise, and the noise doesn't repeat. In our own study of 5,720 configurations, over half of the ones that looked attractive in-sample gave up 50%+ of their risk-adjusted edge on data they had never seen.

The other common causes are mechanical: unrealistic fill assumptions (no fees, no slippage, fills at prices you couldn't get), lookahead bias (using information not available at decision time), and survivorship bias in the asset universe. A backtest engine has to get these right before validation even matters.

The fix is procedural, not clever: hold out data the strategy never sees, count every configuration you tried, and demand the result survives out-of-sample before risking money.