Signature Verticals

Vertical 2 — Forecast Accuracy Track Record

5 min read

What it is

Every AI forecast we publish is logged with its prediction horizon. When the horizon arrives, a worker pulls real OHLCV from Binance and resolves the outcome — refusing if the candle is more than 90 minutes off-horizon. Accuracy is reported with Wilson 95% confidence intervals on real outcomes only; neutral moves under 0.1% are excluded so we never claim credit for noise. Sample size is always visible.

How to think about it

Every other forecasting product on the market reports point estimates without confidence intervals. When their model was right twice and wrong once on n=3, the dashboard says "67% accuracy" and you'd never know it's based on three calls. We refuse to do that.

Step-by-step

  1. Visit Intelligence → Forecast Accuracy

    See global accuracy with Wilson CI, plus per-symbol breakdown.

  2. Read the sample size

    A green badge with n=3 should be ignored; the lower bound of the CI on n=3 is brutally low. Filters above n=20 are where the data starts to mean something.

  3. Check the recent calls

    The "Recent" tab shows individual forecasts with how each one resolved. Disagreements are visible.

  4. Public verification

    Anyone can read the global stats at /api/forecast-accuracy/summary without an account — for journalists, regulators, or anyone validating our claims.

Tips & pitfalls

  • The Wilson lower bound is what de-ranks small samples. A 100% accuracy on n=2 has a Wilson lower bound of ~16% — the math reflects the actual sample size.
  • Neutral moves (<0.1%) are excluded so a forecast of "BTC will go up" doesn't get credited when BTC moves 0.05%.
  • If a candle is more than 90 min off-horizon (e.g. exchange downtime), the forecast is left unresolved — never auto-resolved with stale data.