Intelligence

Forecast Accuracy

4 min read

What it is

Public, verifiable track record of every probability the haythix forecast engine emits. Each forecast gets logged at capture time, then resolved against real OHLCV at horizon by a worker. Wilson 95% confidence intervals are shown so small samples don't mislead. Anyone can hit the public endpoint to verify the numbers independently — no marketing claims.

How to think about it

Most "AI accuracy" pages on competitor platforms cherry-pick or backfill. We log forecasts BEFORE the outcome is known, then resolve against real data. The page surfaces honest stats including the cases where sample size is too small to draw conclusions.

Step-by-step

  1. Pick a window (7d / 30d / 90d / 1 year)

    Top-right selector. Shorter windows = fresher data but smaller sample. 30 days is the default — usually enough samples to be statistically meaningful while staying recent.

  2. 3 horizon tiles (1h / 4h / 24h accuracy %)

    Global accuracy across all symbols at each forecast horizon. Each tile shows the headline % + sample size (decided/total) + Wilson 95% CI band. Color: green ≥60%, blue 50-60%, amber 40-50%, red <40%, amber overlay when sample <10 (too small to trust).

  3. Per-asset breakdown table

    Same accuracy + sample stats broken down per symbol. Find your favorite pair and check how the engine has performed specifically on it — different symbols can have very different accuracy.

  4. Recent Forecasts table

    Last 7 days of individual forecasts and their outcomes. Each row: timestamp, symbol, horizon, predicted direction, predicted probability, actual outcome (hit / miss / neutral), resolution price. Verify the page isn't hiding individual misses.

  5. Public endpoint

    Bottom of page shows the public URL — anyone (no account, no auth) can hit GET /api/forecast-accuracy/summary?days={N} to verify the numbers. Used by journalists, regulators, and skeptical users.

Tips & pitfalls

  • Wilson 95% CI matters more than the headline accuracy when sample is small. 100% accuracy on n=2 has a Wilson lower bound of ~16% — the math is honest about small samples.
  • Neutrals (<0.1% move in either direction) are excluded from the accuracy calculation. They're not "right" or "wrong" — there was no directional move to predict.
  • If the resolver can't fetch real OHLCV at horizon (e.g., exchange outage), the forecast stays unresolved rather than auto-marked. Honest behavior.
  • Per-asset accuracy can vary widely. The engine may be 70% on BTC/USDT and 45% on a thinly-traded altcoin. Check your specific symbols before trusting the headline.
  • The "track record is just starting" empty state is the right answer when the window has no resolved forecasts. Wait or expand the window.