Most retail '90% accuracy' models are overfit mirages. Proper backtesting means walk-forward: train on window one, test on window two, roll forward, repeat. Never train on the future.
Leakage killers include using the same-day close as a label with same-day features, not lagging variables, and including the traded bar in the window. The overfitting checklist: feature count below samples divided by ten; out-of-sample Sharpe above half the in-sample Sharpe; stable across regimes; no look-ahead; costs modeled.
Q: Why is walk-forward better than a train/test split?
A: A single random split can accidentally leak future patterns; walk-forward tests on strictly future data repeatedly.
Q: What is the overfitting checklist?
A: Feature count control, out-of-sample Sharpe retention, regime stability, no look-ahead, costs modeled.
Q: Why publish methodology not just PnL?
A: Methodology is reproducible and credible; a PnL screenshot alone is easy to fake or overfit.
Q: How many folds are enough?
A: Enough to cover multiple regimes — at least several non-overlapping out-of-sample windows across a year or more.
📞 Free Nifty Options AI help & resources — WhatsApp: 9169650895
By Shakti Tiwari · Options AI research pillar. NISM XII certified. Educational only, not investment advice; verify before acting.