Forward test · collecting evidence
Predictions are saved before play, including estimates that fail the shortlist checks. Each selection uses a simulated 1-unit stake at its frozen quote. This is a public model experiment, not a record of your personal bets.
Loading paper ledger…
Probability accuracy
Shots losses deduplicate event/player/line; game losses deduplicate game/market/handicap. Opposite selections are not counted twice. Lower Brier loss is better. Shots calibration uses the over outcome. Game calibration uses home wins/covers, or all three regulation outcomes. It does not establish profitability.
Calibration · all frozen predictions
Frozen selections
Newest 100 selections shown. Passed checks means at capture time. Multiple lines on the same player and opposite sides are correlated alternatives, not independent bets.
Sampling and grading rules
Hourly checks aim to capture the earliest available regular-season games 15–180 minutes before the earlier of the NHL and provider start times. Maximum two attempted games per Eastern calendar day, with an observed 100-credit reserve. This intentionally does not cover every game. Scheduled runs can be delayed; late games are skipped, never backfilled after start.
Best-price selections for shots, moneyline, puck line and regulation three-way are frozen when offered. Game markets add up to 3 provider credits per sampled game; the two-game daily cap and observed 100-credit reserve remain. Missing or ambiguous player data is counted as an exclusion. Quotes, probabilities, model version, source commit and shortlist decisions are preserved with a SHA-256 digest and Git history. Result corrections can update grades without rewriting predictions.
Only official final games are graded. Moneyline/puck line include overtime and the shootout winner goal. Three-way uses the verified 60-minute result; unknown regulation metadata stays pending. For shots, no appearance in the official player log is a paper void after 24 hours. Other settlement differences, postponements and unresolved data remain pending. Corrections are checked daily for 14 days. Pushes return the simulated stake; ROI includes stakes on wins, losses and pushes, and excludes voids and pending selections.
All-vs-shortlist ROI is descriptive. Small samples, correlated selections and multiple testing can exaggerate performance. No bankroll is connected and no wager is submitted.