Backtesting Discipline

Core rules (non-negotiable)

1. No lookahead bias. A signal at bar t may only use data available at close of t (or earlier). Never use t+1 close, future highs/lows, or full-series indicators recomputed on the whole dataset. Shift signals by one bar before applying to returns.
2. No survivorship bias. Include delisted/dead pairs in the universe for the period tested.
3. Realistic costs. Binance spot taker fee = 0.10% per side (0.075% with BNB). Futures taker ≈ 0.04%. Always model slippage (≥ 1 tick, more for low-liquidity alts / large size). A strategy that only works at zero cost is not a strategy.
4. Point-in-time data. Funding rates, OI, on-chain metrics must be as-reported at the time, not revised.

Validation

• Out-of-sample / walk-forward — never report only in-sample. See [[strategy-optimization]].
• Regime split — report performance separately across bull / bear / chop. A strategy profitable only in one regime must say so.
• Multiple-testing awareness — if you tried N parameter sets, the best one's Sharpe is inflated. Deflate or hold out.

Metrics to report (always together)

• Sharpe, Sortino, Calmar
• Max drawdown (% and duration) — see [[risk-management]]
• Win rate, profit factor, avg win/loss, expectancy
• Trade count (low N → not significant)
• Equity curve + per-regime breakdown

Red flags (reject the result)

• Sharpe > 3 on daily data → almost always a bug or leak.
• Smooth equity curve with no drawdowns → lookahead.
• Performance collapses out-of-sample → overfit.
• Results depend on a single trade or single month.