Jack Henry L730 Anderson#39

Посмотреть перевод

What is data bias and overfitting in backtesting—and how do you avoid them?

Data bias includes look-ahead (using info unavailable at the time), survivorship (ignoring delisted assets), and selection bias (cherry-picking periods/instruments). Overfitting occurs when a model memorizes past noise—performing great in backtests but failing live. Prevention toolkit: strict train/validation/test splits; walk-forward analysis; realistic costs/slippage; and parameter parsimony (fewer knobs). Use robust stats: out-of-sample Sharpe, max DD, and turnover-adjusted returns. Add noise to inputs, run Monte Carlo with shuffled trade sequences, and stress test with volatility spikes/gaps. Keep a research diary to avoid “p-hacking.” Finally, start live with tiny size (“probation”) and scale only after a statistically meaningful sample. The goal isn’t perfect backtests—it’s durable edge under messy, changing markets.