Every stage is intentionally separated so later information cannot leak backward into an earlier forecast.
Collect raw meteorological guidance for the exact station, threshold, and scoring window.
Convert evidence into an event probability and confidence before the outcome is known.
Write the official immutable snapshot. Missed issuance stays missing rather than being backfilled.
Observe Kalshi and Polymarket as separate market signals with contract and execution checks.
Score the frozen probability against matured station reality, then keep the result permanently.
These charts are generated from the same persisted production evidence used by the dashboard. Early samples are shown as early samples—not dressed up as certainty.
This separation is one of PRECIPIQ’s core trust boundaries.
PRECIPIQ is designed to make its own mistakes visible and its evidence reconstructable.
Once the official sample is issued, later models, radar, prices, or outcomes cannot rewrite it.
Accuracy and Brier Score sit together because a confident probability should be judged more harshly when it is wrong.
Prediction time, probability, result, evidence source, and misses remain visible after settlement.
A market is useful only when its station, threshold, side, timing, and settlement rule are comparable.
Stale quotes, missing evidence, thin fills, and uncertain comparisons block action instead of being guessed through.
Experimental layers stay shadow-only until out-of-sample evidence shows that they improve decisions.
City history is useful context, but it is deliberately constrained so a hot streak cannot overpower meteorology or execution safeguards.
Tomorrow-specific city performance is shrunk toward the global baseline and gated by sample size. Historical performance cannot make a bad market actionable.
The current Reliability layer remains observational until out-of-sample review proves that using it improves decisions without damaging calibration.
The fastest way to make a forecasting system look smart is to weaken the audit trail. PRECIPIQ goes the other direction.
If an official issuance is missed, newer information is not inserted into the scored historical sample.
Execution freshness and fillability matter. A beautiful theoretical edge with unusable pricing is not a trade.
Kalshi and Polymarket are evaluated independently. Differences are explained, not averaged away.
Wrong predictions remain in history and continue affecting accuracy, calibration, and Brier Score.
Open the live dashboard, inspect the frozen forecast, then follow the receipts after reality arrives.