How a PropBetEdge forecast is made
1 · Model the contract, not the headline
Every venue contract is normalized to its exact resolution semantics: station or series, observation window (for weather, the NWS climate day in local standard time), comparator, threshold, units, resolution authority and exceptions. If the rules text and the venue's structured fields disagree, or the settlement source is unknown or commercial, the contract is held with a machine-readable reason (HOLD_RESOLUTION_AMBIGUOUS, UNMODELABLE, COMMERCIAL_SETTLEMENT_SOURCE) and receives no probability.
2 · The market never enters the model
Kalshi prices, order books, volume and price history are stored as a benchmark only. Feature vectors are built from allowlisted official/public source classes; a code guard and a database constraint reject any market-derived key, and a regression test shows PBE probabilities and feature hashes are unchanged when every market price is shifted.
3 · Point-in-time data
Each input carries the time it became available. Weather guidance runs are used only five hours after their cycle; rates use the official Daily Treasury Par Yield Curve as published; macro inputs use never-revised daily series. Backtests apply the same rules and are evaluated on chronological holdouts.
4 · Immutable snapshots and fixed scoring roles
A forecast is written once with its model version, feature snapshot, data cutoff, evidence and the market price observed at that moment. A later run creates a new snapshot; nothing is rewritten. The snapshots that count for scoring are fixed by rule before the outcome: FIRST_PUBLISHED, T_MINUS_24H (only if a forecast existed 24 hours before the reference time) and FINAL_PRE_RESOLUTION.
5 · Resolution and scoring
The venue settlement and an independent official value (the NWS Daily Climate Report for the exact station, the Treasury par curve, the Federal Reserve target range) are stored separately and compared. PBE and the market are scored on the same snapshot with Brier score and log loss. Model skill is not claimed until enough contracts have resolved.
6 · Model states
MONITORING (market shown, no model) → BACKTESTING → SHADOW (forecasts stored, never published) → RESEARCH (published, not validated) → VALIDATED → OFFICIAL. See the research board.
7 · Access
This methodology, the model registry and the aggregate scoring record are public. The forecasts themselves — every PBE probability, the model-vs-market comparison, the evidence ledger, forecast history and the full intelligence desk — are PropBetEdge Predictions, included with PropBetEdge All Access ($29/month).
8 · Partner relationships
PropBetEdge participates in Kalshi's referral program for Kalshi Perpetuals. Some pages show a separate, clearly labelled Kalshi Perpetuals offer. PropBetEdge may receive compensation from qualifying Kalshi referrals; offer eligibility and terms are determined by Kalshi. That offer is kept apart from market links: "Open on Kalshi" always goes straight to the contract shown, with nothing added. Partner compensation does not affect PropBetEdge model probabilities, market comparisons, rankings, editorial conclusions, or research, and Kalshi data remains a benchmark that never enters a PBE model.