Signal
The model sees enough edge after costs and the risk gates are clear.
Loading
Preparing this view
Retrieving the current information and preserving your place. No action is required.
Loading current information
Forecast AlphaForecast AlphaLoading
Retrieving the current information and preserving your place. No action is required.
Methodology
Forecast Alpha is built around a simple standard: a probability gap is not enough. A useful prediction-market read must show the model estimate, market price, costs, risk gates, and proof trail before it earns a decision label.
Research only. Not financial advice. Prediction markets involve risk including total loss.
Decision pipeline
Pull active contracts, prices, order-book context, metadata, settlement timing, and category-specific features.
Compare market-implied probability with a model estimate built from event context, crypto features, historical outcomes, and confidence signals.
Convert raw disagreement into expected value after modeled fees, spread, slippage, timing, and execution constraints.
Block weak reads when liquidity, ambiguity, stale data, risk, settlement wording, or confidence does not clear the threshold.
Persist model reads before resolution so Brier score, calibration, and beat-rate calculations can be audited later.
Decision gates
The model sees enough edge after costs and the risk gates are clear.
The market is interesting, but the edge, confidence, or timing is not strong enough yet.
A blocking condition makes the market unsuitable despite any apparent probability gap.
Performance metrics
The public product emphasizes metrics that can be checked after resolution. Wins matter, but so do the probability quality, the cost-adjusted decision, and the cases where the system correctly refuses to act.
Probability accuracy. Lower is better; 0 is perfect and 0.25 is chance for a 50/50 forecast.
Whether events forecast at a given probability happen at roughly that frequency.
How often the model's probability scores better than the market probability on the same resolved row.
Paper or live P&L quality: average gain relative to average loss after costs.
How often the system refuses to act because a gate blocks the read.
The exact model family used for a prediction so performance can be compared across versions.
The five-gate bar
At least 100 resolved, out-of-sample predictions scored on real (non-demo) markets, accrued over at least 30 days of forward shadow operation — no backfills, no retroactive selection.
Better log loss than the current champion on the identical resolved set, with no category degrading by more than 0.05 Brier — a model can't buy its headline in one category by getting worse in another.
The per-prediction Brier improvement is positive with a 90% paired bootstrap confidence interval (2,000 iterations) that excludes zero.
Calibration error no worse than the champion's — stated probabilities have to mean what they say, not just rank outcomes.
Cross-fold weight stability (no sign flips on significant factors) and no degradation in NO_TRADE quality — the refusal gates must keep skipping bad bets.
Pre-gate work runs as pre-registered experiments on the track record page — hypothesis, thresholds, and required sample size fixed and dated before the data, pass or fail shown either way.
Verification
Forecasts are useful only if they are recorded before outcomes are known. Forecast Alpha persists model reads with timestamps, market probabilities, model versions, and outcomes once available. That lets users inspect not just whether a market was called correctly, but whether the stated probability was well calibrated over time.
FAQ
No-trade decisions show where apparent probability gaps fail cost, liquidity, confidence, timing, ambiguity, or data-quality gates.
Brier score and calibration are the primary probability-quality metrics. Paper P&L and win rate are useful, but they are secondary to whether probabilities match outcomes over time.
No. Forecast Alpha is a research and decision-support terminal. It does not guarantee outcomes or provide financial advice.