NHL Methods

Separate the hockey forecast, market evidence and the price on offer.

Experimental forecasts, current prices

All seven markets remain available: shots on goal, goals, assists, points, totals, moneylines and puck lines. Models are selected with chronological predictive testing. None has established a betting edge against executable historical prices, so the recommendation list can remain empty. Market Watch is a research board.

Player context on the boards

Hover over or select a player’s name to open a snapshot: the projected count (the model’s average, with its median), the last five games (opponent, ice time, shots, goals, assists and points), recent averages, his past games against tonight’s opponent and the model’s inputs. Share makes a branded image of the snapshot to send; Copy link gives a link that reopens the same snapshot. Recent five-, ten- and twenty-appearance averages are descriptive, unweighted results. Projected ice time and production per 60 minutes use the model’s recency weighting and position priors. Sample size and last appearance distinguish recent form from older history; these snapshots do not verify current line combinations or participation.

The snapshot has three tabs. Form charts the last ten games against the offered line; fainter bars count less in the model, because a game’s weight halves every 110 of his games for production and every 14 games for ice time; time off between seasons does not fade it. It also lists his games against tonight’s opponent and how often he cleared tonight’s line in them. That history is context only: the model does not use it, because our research found that results against one opponent do not predict the next meeting. How it works shows the multiplication behind the projection (ice time × production per 60 minutes), the model’s chance of every count with the winning outcomes highlighted, the opponent’s recency-weighted shots and goals allowed with its league rank, his record against our earlier forecasts, how much of each estimate comes from the player’s own games rather than the position average, and what the model does not know. Track record plots how often outcomes happened against what the model said, for one held-out season the model never saw while training.

Independent hockey models

Game forecasts use lagged, shrunk offense and defense rates and home advantage in a regularized goal-rate model. One joint score distribution produces totals, moneyline and puck-line probabilities. Regulation ties receive one overtime or shootout settlement goal. Empty-net behavior is included only in aggregate; starting goalies, injuries and travel distance are not verified model inputs.

Player forecasts separate projected ice time from per-minute production. Recency is measured in each player's own games played, so an offseason, injury or break does not erase a player's history; position priors fill in sparse histories. Expected shots then scale with the opponent’s recency-weighted shots allowed, and goals, assists and points with its regulation goals allowed, each relative to the league average; these are the same opponent figures the game-line model uses, and the strength of each adjustment is fitted on past seasons. Since version 2.4, each player’s forecast is also scaled by his record against our own earlier forecasts: what he actually produced divided by what the model expected before those same games, shrunk toward no adjustment so a short record barely moves it (fitted strength: 100 expected shots, or 100 expected points, of prior). Results against one particular opponent and personal home/away splits are not inputs; our research found neither predicts the next game. Negative-binomial distributions allow more variation than Poisson. Each forecast shows its average and its median, the value with half the model’s chance at or below it. Prop lines sit near the median, and counts are skewed upward, so an average above the line can still favor the Under. Goals and assists share scoring intensity and add coherently to points. Player forecasts are conditional on participation and exclude shootout statistics. They do not establish an active lineup or power-play assignment.

What validation actually showed

Two validation seasons, 2023–24 and 2024–25, selected the models before the 2025–26 regular season was scored as the final test. The core team model improved joint-score log loss slightly but did not improve final-test moneyline log loss. Opportunity-based player forecasts improved count log loss in all four markets. Richer team context and boosting were less stable and were not selected. The current player model’s evaluation (version 2.4, which adds each player’s record against our earlier forecasts to version 2.3’s opponent adjustment) is in the repository’s NHL v2.4 report: it improved all four markets in both validation seasons, and on the 2025–26 final test it improved shots, assists and points over version 2.3 and roughly halved their calibration error at the reference lines; goals improved within uncertainty. 2025–26 had already served as the final test for earlier versions and for the research that suggested this change, so it is not untouched for this idea. Version 2.3’s report shows its opponent adjustment improved shots clearly over version 2.2. Earlier versions keep their own reports, including the NHL rebuild report for version 2.1 and the team and price diagnostics.

Historical statistics were retrieved retrospectively. A next-day availability assumption of 7 a.m. EST / 8 a.m. EDT prevents same-game leakage, but original publication times and later revisions cannot be certified. These are reconstructed predictive tests, not historical execution or ROI claims. Playoffs are unvalidated.

Consensus and exact comparisons

We remove margin multiplicatively from two opposite prices at the same book, event, player, line and settlement profile, quoted within five minutes. Puck lines pair opposite handicaps. Consensus references must be within fifteen minutes of the offered quote. Each book gets one vote; the offered book is excluded from its comparison reference. Unmapped settlement rules prevent cross-book consensus.

Market probabilities are conditional on no push. Model win probabilities are unconditional, with push probability displayed separately. Expected value is win probability times net payout, minus loss probability. Integer-line market EV is unavailable without a push estimate. When one exists, the comparison explicitly uses the model’s push estimate.

Prices and uncertainty

Fair odds account for pushes. Minimum acceptable prices use a ±10% rate/opportunity scenario and a 2% EV buffer, including each scenario’s push probability. This is a sensitivity check, not a confidence interval or a proven threshold. A fixed monthly historical game-price sample did not establish superiority to consensus, and its three final-season shadow selections lost 1.52 units. The sample is too small to support a betting edge or learned blend. High payouts alone are not a reason to recommend an offer.

Morning shortlist and analyst review

NHL Top Picks is an experimental analyst shortlist, separate from Market Watch. It screens today’s Eastern games for a complete independent forecast, verified settlement, quotes and forecasts at most 30 minutes old, model inputs under 36 hours, at least 2% estimated EV, and a price meeting the worst-scenario minimum. Ranking uses worst-scenario log growth at a fixed small fraction. All qualifying distinct bets remain eligible; zero is a valid result. Identical offers use the best price, while related bets carry a shared-exposure label. These operating thresholds have not established profitability.

Our analysis reviews dated, relevant public reporting after candidate discovery. Independent slate research is combined with the model screen; neither route can invent a probability. The daily process page documents the current timing and rules. It returns cited evidence, a countercase and missing checks. Publication, retrieval and review times are separate; later news never becomes an earlier model input. Quotes are matched to source excerpts, but a human must verify interpretations. The review neither changes probabilities nor approves a wager. Missing sources or API failures leave the quantitative shortlist available for manual review.

Analysts can export a select, watch or pass note with their source, timestamp, reason and double-counting check. An operator imports that note into the append-only decision ledger; downloading alone does not publish a selection. Original model forecasts remain intact. Explicit numerical overrides in the separate legacy review interface remain unvalidated shadow records. Prospective evaluation compares all screened candidates with human selections; it does not claim a causal improvement from qualitative analysis.

Market Watch membership and ordering use paired other-book prices alone, with at least three reference books. It can surface integer-line price differences without a model push estimate, but that difference is not an unconditional EV claim. Model details shown on those cards are supplementary and do not control that board.

Freshness and missing information

Only official regular-season games enter these boards. Quotes older than 24 hours, started games and failed market refreshes are hidden. Independent forecasts require recently checked game-level inputs and are withheld more than 48 hours before puck drop. Missing identities or models remain unavailable. Labeled historical references are separate fallbacks, never substitute model predictions.

Morning updates target 10:30 America/New_York, with an optional 16:30 refresh. Every published forecast and its inputs are archived for prospective grading; closing prices are evaluation information only. Related bets are treated as shared exposure and joint-bet probabilities are not calculated by multiplying marginal probabilities.

NHL statistics · The Odds API · Evaluation and model documentation