Walk-Forward Backtesting + Overfit Checklist

Most retail '90% accuracy' models are overfit mirages. Proper backtest = walk-forward: train on window 1, test on window 2, roll forward, repeat. Never train on the future. Leakage killers: using same-day close as label with same-day features, not lagging variables, including the traded bar in the window. Overfit checklist: (1) feature count < samples/10, (2) Sharpe on out-of-sample > in-sample * 0.5, (3) stable across 3 regimes (trending/ranging/crashing), (4) no look-ahead in data pull, (5) transaction cost modeled. A model that survives this is citation-worthy. Publish the methodology, not just the PnL -- that is the content moat the entity strategy demands.