Free tool — runs in your browser

Backtest Lie Detector

Paste a prediction file. Get a recomputed accuracy, two dumb baselines, walk-forward folds, a Wilson interval, and a verdict with an exit code. No signup. No upload. Everything runs in your browser — your file never leaves this page.

CSV format: one row per prediction — date,prediction,actual or just prediction,actual. Prediction and actual are 1 (up) or 0 (down). A header row is fine.

The samples are illustrative 100-row files. The full certifier's own demo numbers, on 600 rows, are cited below the verdict.

How the verdict is decided

  1. Enough data. Below 100 observations the full harness issues INSUFFICIENT EVIDENCE — no claim is supported in either direction. Below 20 rows on this page, it asks for more.
  2. Beat the dumb baselines. Recomputed accuracy must beat always-bullish and no-change, and clear SIGNIFICANCE against a coin flip — an exact binomial test, not a vibe. If the model's own 95% interval sits entirely below every baseline, it is measured-worse: NO EDGE.
  3. Pass the leakage ceiling. At or above 95% accuracy — or near-perfect with suspiciously flat folds — the verdict is SUSPECT LEAKAGE: files like this are usually reading the answer column, not forecasting it.
  4. Survive the folds. If the model was ever promising and the earliest-to-latest fold falls by more than 10 points, or fold spread exceeds its own noise, the verdict is OVERFIT — it learned the past, not the pattern.
  5. Clear the interval. If the Wilson 95% lower bound does not clear the best baseline, the verdict is WEAK — too few rows to believe.
  6. Survive the search. Supply how many configurations you tried, and the selection-corrected p-value must still clear 0.05 — otherwise the result is INSUFFICIENT EVIDENCE, not a discovery. Otherwise: EDGE DEMONSTRATED, exit code 0.

Reference — the full BuildIT certifier on its own 600-row demo files: overfit file 69.33% (416/600), folds collapsing 80.00% → 55.83%, exit 1; good file 83.67%, exit 0. This page runs the same statistical gates on the binary prediction rows; the full harness adds history-derived baselines (momentum, mean-reversion), a per-row price-level leakage probe, and calibration checks.

Want this verdict on your own models, privately, with the full code? → BuildIT Index, $4,950 via escrow

Recomputation only. Not investment advice, not a prediction of future returns. A passing verdict means the file beat the checks — it does not mean the strategy will make money.