Forward Test

Paired intermission comparison

394 identical intermissions · 198 games. Both models use the same clock, shot attempts, market quotes and edge threshold. Predictions include every eligible intermission; bet counts can differ when only one model finds enough edge.

Paired · same intermissions · 2026-09-29 – 2026-10-27
paired-v1-a20cdf2ec604

Theoretical flat-bet P&L

1u of total spend per qualifying bet, including fees. Settled bets only; all orders assumed fully filled. No compounding.

2026-10-06 · OTT @ DET P2
2026-10-13 · VAN @ NYI P2
2026-10-28 · PHI @ WSH P1
2026-10-28 · PHI @ WSH P2

2026-09-29 · CHI @ VGK P2
2026-10-06 · OTT @ DET P2
2026-10-06 · FLA @ LAK P2
2026-10-13 · VAN @ NYI P2
2026-10-13 · PIT @ CHI P2
2026-10-13 · MIN @ WPG P2
2026-10-13 · COL @ DAL P1
2026-10-13 · COL @ DAL P2
2026-10-13 · VGK @ NSH P1
2026-10-13 · VGK @ NSH P2
2026-10-13 · TOR @ UTA P1
2026-10-13 · TOR @ UTA P2
2026-10-13 · CGY @ ANA P1
2026-10-13 · CGY @ ANA P2
2026-10-13 · EDM @ LAK P1
2026-10-13 · EDM @ LAK P2
2026-10-13 · BOS @ SJS P1
2026-10-13 · BOS @ SJS P2

duplicate predictions: 132

Predictions come from pressure_signal_log.jsonl (intermission decision, about 3 minutes into the break) and live_alerts.jsonl (ELO Raw / Predictive / Full at the initial alert). Each uses its own logged market probability. ELO and intermission snapshots may have different timestamps; these are separate matched comparisons.

We keep the earliest valid game / period snapshot in each model family, including across monitor restarts. Paired snapshots record both raw-pressure and team-style predictions together; both are retained from the same earliest valid snapshot. The original-live view includes historical active-only records without recreating missing forecasts. The walk-forward view rebuilds both intermission models using prior-season slopes and prior-date team baselines on the same breaks; it is a retrospective backtest, not a record of what ran live. Source breakdowns remain available for inspection. Legacy snapshots lack an exact model fingerprint; changes within an unversioned cohort cannot be identified retrospectively. Context-only probabilities are excluded from the primary order model.

Only official NHL FINAL or OFF scores, joined on game ID and both team identities, settle outcomes. Missing results stay unscored. Recorded next-period starts exclude late decisions when available. Otherwise timing relies on the immutable intermission log; old records have no independent final-time stamp. Dates are NHL game dates, falling back to the original timestamp in Eastern time for legacy ELO alerts.

Log loss = −log(probability assigned to the winner), using natural logs and clipping to [0.000001, 0.999999]. Brier = mean squared error. Delta = model loss − market loss on identical settled snapshots. Calibration averages home forecasts versus home win frequency. Intervals resample entire games with both periods together, 1,000 paired bootstrap draws; fewer than 10 games have no interval. A small sample does not establish a durable edge.

Execution prices use the recorded outcome side: home maker = home bid, home taker = home ask; away maker = 1 − home ask, away taker = 1 − home bid. The logged order price is authoritative for maker pricing. Missing/crossed quotes stay missing; no prices are inferred from the midpoint. Each cohort has raw make/take and style make/take rows. Logged act=true signals with both quotes are candidates; each strategy requires (model probability for the chosen side − fee-inclusive entry cost) × 100 ≥ the recorded threshold in percentage points. The quote and fee are per share; fees use the market’s recorded or separately recovered schedule. Supported sports fees are rate × price × (1 − price), with exponent 1 and taker-only fees; missing or unsupported settings are gaps. Makers have zero fees and no assumed rebate. Model and cost log loss score that strategy’s qualifying settled bets; cost loss uses the fee-inclusive break-even price on the purchased outcome. Delta = model loss − cost loss. Forecasts are unchanged but different selected bets can change model loss. Average costs include qualifying pending games. These are recommendation scores, not verified fills or realized ROI.

Theoretical flat-bet P&L uses the same qualifying recommendations as the execution scores and the daily Discord report. Each bet spends one unit including fees, buying 1 / (price + fee) shares. A win earns 1 / (price + fee) − 1 units; a loss loses one unit. ROI = total profit / settled units staked. Pending or missing outcomes contribute neither profit nor settled stake. All fills are assumed, with no compounding, liquidity limits or slippage. One unit represents any fixed dollar stake; multiply P&L units by that stake for dollar profit. Totals follow the selected dates and model/period filters, combining all recorded versions and historical/paired observations. Each loss is averaged over its underlying scored observations; ROI uses total profit divided by total settled stake.

Default logging window starts Sep 29, 2026. Counts cover recorded predictions and expose missing completed-game windows separately. This page reads existing logs; NHL results use a 2-minute server cache for active dates and 6 hours for completed dates. Derived walk-forward files remain separate from immutable live logs. No orders are placed.