What is this?
A live leaderboard measuring how well AI and algorithmic models predict short-horizon price direction in crypto (BTC/ETH) and forex (EUR/USD) markets. Predictions are recorded to a public database before outcomes are known, then scored automatically when each market window resolves.
What exactly is being predicted?
Every model calls the direction of the price move over a fixed 24-hour window: will the asset close higher or lower 24 hours from the moment the call is made? There are no 15-minute or intraday calls here, and no price targets — only up/down plus a confidence.
- When calls are made: once a day, ~12:00 UTC (8am ET), for all assets at once.
- When they settle:
- Crypto & forex — exactly 24h after the call (price at call time vs price 24h later).
- Stocks & ETFs — at the next 4pm ET close after the window (a Friday call settles Monday's close; weekends skip).
- Flat settle (~9% of windows) — if the price is unchanged at settlement, the call scores wrong either way.
- Scoring cadence: the resolver runs daily, so a resolved row's timestamp can trail the settle time by up to a day — the Predictions tab shows both timestamps.
Methodology
- Directions, not prices. Each model emits up/down calls with a confidence; scoring compares against realized direction at the timeframe horizon.
- Brier score = (confidence − outcome)² averaged over resolved predictions; lower is better. 0.25 = coin flip at full confidence.
- Accuracy = share of resolved calls matching realized direction.
- P&L = simulated unit stake per call at the model's confidence.
- Random baseline is included as the floor any real signal must beat.
- "How wrong / how right." Direction alone doesn't say how big the miss was. The Predictions tab shows the entry → settle price and % change for every call: a miss by +0.02% is a near-coin-flip, a miss by −4% was confidently wrong. Being "right" means right at settlement (24h / next close) — the path within the window doesn't matter.
Data
All predictions and results are stored in a public Supabase (PostgREST) database and synced from the scoring harness on a schedule. This page computes every statistic client-side from the raw rows — refresh to re-score.