Forecast AlphaForecast AlphaPublic proof
Forecast Alpha audit ledger
Calibration trace
Predicted probability against realized frequency.
How to audit this
Forecast probability, market probability, model version, and timestamp are written before resolution.
The row is graded only after the market resolves, so pending rows stay visibly pending.
Brier contribution compares the model against the market-implied probability on the same contract.
The trust question is whether probabilities match outcome frequencies over many resolved markets.
Latest resolved
No resolved rows in the current filter. Switch status or category to inspect scored forecasts.
Proof slices
| Category | n | Win% | Model B | Mkt B | Beat% |
|---|---|---|---|---|---|
| Crypto | 38 | 68% | 0.086 | 0.118 | 71.0% |
| Macro | 21 | 0% | 0.010 | 0.012 | 47.6% |
| Company | 21 | 24% | 0.022 | 0.028 | 52.4% |
| Fed | 20 | 55% | 0.061 | 0.087 | 65.0% |
| AI Regulation | 11 | 0% | 0.025 | 0.038 | 63.6% |
| CPI / Inflation | 8 | 100% | 0.061 | 0.082 | 62.5% |
| Version | n | Win% | Model B | Mkt B | Beat% |
|---|---|---|---|---|---|
| v1.0.0 | 119 | 42% | 0.050 | 0.069 | 61.3% |
Latest deltas
Record basis
Resolved predictions bucketed by predicted probability vs realized YES frequency over 119 events. A calibrated forecaster sits on the diagonal.
PRO research stack
The public record shows whether the model is calibrated. PRO adds the tools for turning that proof into repeatable market review.
General-engine rows come from model_predictions. Short-term crypto rows come from short_term_crypto_forecasts. In both cases, the displayed forecast was written at prediction time with the model probability, market probability, model version, and timestamp needed to score it later.
General-engine predictions are append-only and enforced at the database level: UPDATE and DELETE statements on it fail with an error, for everyone, including us. A prediction that turned out wrong cannot be edited or removed - it stays on this page forever.
Short-term crypto forecasts keep the prediction fields fixed and fill outcome, Brier, and log-loss fields only after settlement. That keeps the intraday record scoreable while preserving the original forecast probability and timestamp.
Demo-market predictions are simulated and shown only under the clearly-labeled "Simulated history" tab - headline metrics come from live-market predictions only.