V3 — Build 13 / Methodology
The skater-driven NHL Value Lab updates daily. Explore how current-season evidence and established player history inform roster strength and the season ahead. Goaltending is not included in the current predictive signal.
Updated October 19, 2026 at 1:21 a.m. · 100,000 seasons per production seed01
Each part of a player’s game has its own memory. Scoring can benefit from a longer history; shot generation uses a shorter window. Older seasons fade rather than counting equally.
| Metric | Category weight | Historical horizon | Decay / curve | Shrinkage K | Age adjustment |
|---|---|---|---|---|---|
| ScoringGoals / GP | 26.84% | 14 seasons | Exponential · 1.5-season half-life | 20 | Position-specific, partially pooled |
| PlaymakingAssists / GP | 27.10% | 7 seasons | Exponential · 1.25-season half-life | 20 | Position-specific, partially pooled |
| Shot generationShots / GP | 6.42% | 3 seasons | Recent-season factors: 1, 0.5, 0.2 | 20 | Position-specific, partially pooled |
| 5v5 possession5v5 SAT Relative | 8.67% | 11 seasons | Exponential · 1.5-season half-life | 40 | Position-specific, partially pooled |
| Puck managementTakeaways - Giveaways / GP | 4.25% | 13 seasons | Exponential · 1.25-season half-life | 40 | Position-specific, partially pooled |
| Shot blockingBlocks / GP | 6.47% | 14 seasons | Exponential · 1-season half-life | 20 | Position-specific, partially pooled |
| DisciplineNet Penalties / GP | 2.22% | 12 seasons | Exponential · 4-season half-life | 80 | Position-specific, partially pooled |
| Penalty-kill deploymentSH TOI / GP | 8.12% | 14 seasons | Exponential · 0.5-season half-life | 20 | Position-specific, partially pooled |
| Goal-differential proxyPlus-Minus / GP | 9.92% | 12 seasons | Exponential · 1.25-season half-life | 120 | Position-specific, partially pooled |
Player Rating = 50 + 10 × the weighted sum of the nine projected standardized components. Historical shrinkage and forward/defenceman age effects remain frozen. A rating of 50 is the historical reference level; Total TOI / GP is excluded from predictive value; the nine-metric weights and normalization moments are frozen.
02
This matrix shows each season’s share of the recency weights, assuming equal games played. Shrinkage and age adjustments follow separately. A dash means the season is outside that metric’s horizon.
| Metric | 2025–26 | 2024–25 | 2023–24 | 2022–23 | 2021–22 | 2020–21 | 2019–20 | 2018–19 | 2017–18 | 2016–17 | 2015–16 | 2014–15 | 2013–14 | 2012–13 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Scoring | 37.1% | 23.3% | 14.7% | 9.3% | 5.8% | 3.7% | 2.3% | 1.5% | 0.9% | 0.6% | 0.4% | 0.2% | 0.1% | <0.1% |
| Playmaking | 43.5% | 25.0% | 14.3% | 8.2% | 4.7% | 2.7% | 1.6% | — | — | — | — | — | — | — |
| Shot generation | 58.8% | 29.4% | 11.8% | — | — | — | — | — | — | — | — | — | — | — |
| 5v5 possession | 37.2% | 23.5% | 14.8% | 9.3% | 5.9% | 3.7% | 2.3% | 1.5% | 0.9% | 0.6% | 0.4% | — | — | — |
| Puck management | 42.6% | 24.5% | 14.1% | 8.1% | 4.6% | 2.7% | 1.5% | 0.9% | 0.5% | 0.3% | 0.2% | <0.1% | <0.1% | — |
| Shot blocking | 50.0% | 25.0% | 12.5% | 6.3% | 3.1% | 1.6% | 0.8% | 0.4% | 0.2% | <0.1% | <0.1% | <0.1% | <0.1% | <0.1% |
| Discipline | 18.2% | 15.3% | 12.9% | 10.8% | 9.1% | 7.6% | 6.4% | 5.4% | 4.5% | 3.8% | 3.2% | 2.7% | — | — |
| Penalty-kill deployment | 75.0% | 18.8% | 4.7% | 1.2% | 0.3% | <0.1% | <0.1% | <0.1% | <0.1% | <0.1% | <0.1% | <0.1% | <0.1% | <0.1% |
| Goal-differential proxy | 42.6% | 24.5% | 14.1% | 8.1% | 4.6% | 2.7% | 1.5% | 0.9% | 0.5% | 0.3% | 0.2% | <0.1% | — | — |
On smaller screens, swipe or focus the table to explore older seasons.
03
Early results add a little information. As a player plays more games, the current season carries more weight while the historical projection still provides an anchor.
| Games played | Current season influence | Historical prior influence |
|---|---|---|
| 0 | 0.0% | 100.0% |
| 5 | 10.9% | 89.1% |
| 10 | 19.6% | 80.4% |
| 20 | 32.8% | 67.2% |
| 40 | 49.4% | 50.6% |
| 60 | 59.4% | 40.6% |
| 84 | 67.2% | 32.8% |
The NHL schedule here is 84 games. Historical seasons use an 82-game basis, so current games played are converted onto that same basis before blending. This keeps the historical and current evidence comparable. Players can receive a rating as soon as they have one NHL game with complete supporting statistics. Twenty games is not an activation threshold.
GP82 = GP × 82 / 84. Current-season weight = GP82 / (GP82 + 40). Updated component = (1 − weight) × historical prior + weight × current standardized rate. Frozen 2025–26 means and standard deviations preserve the scale. The starting prior follows a fixed order: the active max14 projection, an exact preserved archived max14 projection, a reconstructed max14 projection for a returning veteran with older official NHL history, a K40-shrunk low-sample NHL history, then an empirical forward/defenceman fallback for a true newcomer with at least one NHL game. Zero-career-game newcomers remain NR. Older evidence fades and is shrunk through the same frozen max14 formula; it is not treated as current form.
Known Lab players with approved historical rating inputs receive a rating even without a current-season game or a place on the current-roster snapshot. Injured players retain that rating while remaining excluded from team scoring. A rating does not change affiliation or contract eligibility.
A known Lab player with at least one current NHL game and complete official statistics can receive a rating even when omitted from the current-roster snapshot. The frozen historical seed is used when available; a true newcomer uses the existing position fallback. Zero-game newcomers remain NR, and older established players without an approved seed remain review cases. This does not change a player’s validated team or eligibility. Eligible rated players compete for the selected lineup, with the others retained as depth. Promoted ratings restart from their frozen seed before applying current-season evidence once.
Rating and team eligibility are separate. An unsigned free agent or PTO player may keep a rating but does not count toward a team until validated eligibility evidence allows it. Five official NHL report families—summary, realtime, penalties, scoringRates, and timeonice—must reconcile by player ID, games, season, and units. Pagination is exhausted; duplicates, conflicts, missing components, and partial coverage block the entire update. Official roster and eligibility evidence is checked separately.
The organizational pool includes verified signed depth beyond the daily roster. Official prospect and compensation records help identify gaps, but listings alone do not prove a current NHL contract. Current specialist contract tables can corroborate official signing and team records, including entry-level contract slides. Reviewed identity and contract evidence establish team eligibility; supported NHL history establishes the rating separately. Zero-game newcomers remain NR. Unresolved status, unavailable discovery sources, and contract reviews due are reported for further review.
The daily updater also checks official signings and trades for players missing from the pool, including off-roster depth. It verifies NHL identity, current contract coverage, and transfers before admission, then applies the existing rating and availability rules. Newcomers without NHL games remain NR; missing or conflicting evidence stays queued for review.
04
The available lineup feeds the season forecast. Completed games stay real; only the remaining schedule is simulated.
Current Lineup ranks each team’s best available 12 forwards and 6 defencemen. Full Strength reselects the same number from all rated, eligible organizational players, assuming injured, suspended, and personal-leave players are available. Both use the same player ratings, replacement reference, and team model; neither sums the whole roster. Ranks use unrounded strength ratings, with team abbreviation breaking exact ties, and remain league-wide when filtered. Current Lineup is the default sort. The team drawer shows both ratings, while its profile and roster remain the current available view. Reviewed ended-career and not-expected-to-resume cases remain excluded from both lenses while their contracts continue. A verified current-season NHL appearance or explicit official comeback announcement automatically removes the career hold. Separate injuries, suspensions, and personal leave still follow ordinary availability rules. Estimated dates, practice, and roster listings cannot remove the hold. Full Strength is a hypothetical lineup baseline, not another Points Outlook or an estimated return date. The scalar lineup selector and multi-component team fit can occasionally produce a lower Full Strength rating; these values retain the existing model rather than forcing an increase.
| Input | Purpose |
|---|---|
| Best projected 12 forwards / 6 defencemen | Selects the best currently available eligible skaters. |
| Scalar player quality | Captures the overall player-value signal. |
| Nine component averages | Preserves how the lineup creates value. |
| Ridge regularization α = 100 | Reduces overfitting in small historical team samples. |
| Completed games | Fixed as actual results, including official standings counters. |
| Remaining games | Simulated using the frozen game model and season uncertainty. |
Value Lab Rank orders teams by unrounded Team Rating, highest first. Exact rating ties use team abbreviation for stable ordering. Banked team results do not directly enter this roster rank; current player statistics still update player value. Forecast rank remains ordered by unrounded mean season points and the existing standings-counter tiebreaks. Points Outlook spans the mean team forecast and a smoothly calibrated standings-scale total anchored to the simulated median first- and last-place totals; it is a presentation range, not a confidence interval. Completed points remain fixed, so the projected portion diminishes as the season progresses and both ends equal the team's actual points after 84 games. Playoff Chance means qualifying, not winning a first-round series.
Confirmed injured, suspended, and personal-leave players keep their ratings, statistics, and organizational records in separate Players-drawer sections. They do not contribute to team scoring while unavailable; eligible replacements enter the model lineup. Healthy scratches remain eligible. Future simulations use current ESPN estimated return dates, with CBS as a fallback, as provisional working dates—not promises of clearance. Dated estimates also apply to illness when provided. Verified day-to-day and week-to-week labels supply empirical return distributions when no date is available. Missing, expired, or after-season estimates retain absence until better evidence arrives; another active absence reason still blocks return. New dates are checked daily even if the injury status is unchanged. Current Lineup and Full Strength ranks describe today’s two lineup views; future returns affect Points Outlook and playoff chances, not today’s designation. Estimated recovery dates, practice reports, and disappearance from an injury feed do not establish clearance; no extra injury penalty is applied.
The 10 inputs use frozen historical standardization and coefficients. Internal Team Strength = 50 + 10 × (Unified input points82 − historical mean) / historical standard deviation. Historical RS and scalar 18S remain research diagnostics, with no public current-model toggle.
05
Every dimension has an NHL rank. Top-five areas are elite; 6th–10th are clear strengths; 11th–22nd are broadly middle of the league; 23rd–27th are clear weaknesses; 28th–32nd are major weaknesses. A balanced team does not need a forced list of strengths or weaknesses.
| Dimension | What it tells you |
|---|---|
| Scoring | Projected goals per game. |
| Playmaking | Projected assists per game. |
| Shot generation | Projected shots per game. |
| 5v5 possession | Projected relative shot-attempt share at five-on-five. |
| Puck management | Projected takeaways minus giveaways per game. |
| Shot blocking | Projected shot blocks per game; one defensive contribution, not complete defensive ability. |
| Discipline | Projected penalties drawn minus penalties taken per game. |
The profile graphic compares independent dimensions with a league-average rank reference. They are not portions of a whole. Penalty-kill deployment and plus-minus remain internal predictive inputs; public profiles show seven dimensions. Descriptive profiles do not establish causal effects.
Both today’s lineup and the frozen end-of-2025–26 roster use the same V3 Unified mapping. The baseline uses the same frozen B2 nine-metric ratings and Unified mapping, including reconstructed vectors for departed players. The difference reflects roster, eligibility, lineup, and player-projection updates. It is neither standings improvement nor a transactions-only measure.
The current replacement statistic is the 25th percentile of eligible rated scalar player values, recomputed after a complete update for descriptive depth use. Missing rated lineup slots retain the frozen replacement scalar and neutral component contribution. Reserve quality did not earn a separate causal headline coefficient.
06
Compare models on the same historical test cases. Lower error is better. Unified improved the primary continuous, game-probability, and joint-simulation playoff evidence; scalar 18S retained the best pure-rank result in the final joint simulation.
| Measure | RS | Scalar 18S | Unified |
|---|---|---|---|
| Points RMSE | 13.493 | 13.458 | 13.183 |
| Rank MAE | 6.577 | 6.414 | 6.703 |
| Game Brier | 0.241775 | 0.241875 | 0.241143 |
| Game log loss | 0.676494 | 0.676737 | 0.675116 |
Historical direct team/game scoring from the final selected-model comparison. Each column reports the same test; lower is better for every metric below. Predictive accuracy and independent-seed stability answer different questions.
| Measure | RS scalar | Scalar 18S | Unified |
|---|---|---|---|
| Points RMSE | 13.343 | 13.370 | 13.036 |
| Rank MAE | 6.671 | 6.461 | 6.654 |
| Playoff Brier | 0.200176 | 0.197615 | 0.195031 |
| Playoff log loss | 0.593955 | 0.589476 | 0.571856 |
Historical joint-simulation results from the final production-safe selected models. Lower is better for all four measures. The earlier reserve-burden candidate was rejected on coherence and production-safety grounds and is not part of this comparison.
How to read the result: Scalar 18S retained the best pure rank MAE. Unified improved simulated points accuracy and both playoff-probability scoring measures.
| Evidence | Interpretation |
|---|---|
| Published V3 checkpoints | Evaluate only eligible results after each checkpoint’s publication timestamp. Never rewrite the forecast. |
| Preseason replay | Retrospective comparison only; it was reconstructed after architecture selection. |
This saved forecast used 100,000 production seasons with direct precision checks. New V3 forecasts use one fixed 100,000-season run. The player model and hockey assumptions are unchanged.
| Check | Required standard |
|---|---|
| Production run | 100,000 seasons; no mandatory diagnostic run |
| Mean-points uncertainty margin | ≤ 0.20 points |
| Event-probability uncertainty margin | ≤ 1.0 percentage point |
| Distribution precision | ≤ 1.0 percentage point of CDF error; point-quantile and Points Outlook sensitivity recorded |
| Every simulated season | League points conserved; exactly 16 qualifiers; completed games never stochastic |
Precision checks target 95% simultaneous coverage across 320 checks. Mean intervals use a large-sample normal approximation, probabilities use Wilson intervals, and distribution bands use DKW. These uncertainty margins replace differences between independent runs; their statistical meaning is different. A failed precision check blocks publication. No automatic extension or reroll is allowed. Points Outlook remains a presentation range, not a confidence interval, and numerical precision is not proof of predictive accuracy. Historical forecasts retain their original counts and gates.
07
These ideas were tested or assessed without becoming active forecast coefficients. A zero schedule multiplier performed best. The forecast cannot know future injuries, coaching or role changes, prospect breakouts, or goalie surprises. Architecture selection and calibration remain frozen.
08
Trades rebuild the affected rosters and selected lineups using the same model. The current Trade Simulator defaults to Full Strength, treating all 32 teams as available before and after the move; Current Lineup retains verified absences today and uses the same supported injury-return assumptions on both sides of the hypothetical trade, mapped to each destination’s schedule. Both scenarios preserve banked results. These hypothetical outlooks are separate from the main Lab forecast. The authorized public standard is 500 matched remaining-season simulations for responsive before/after comparisons. This supersedes the earlier 5,000-run planning target and remains separate from the production forecast standard. New V3 forecasts use one 100,000-season run with direct precision checks; archived forecasts retain their original standard. Small probability differences can include sampling noise. Trades never enter the forecast archive or accuracy records.
| Record | What it means |
|---|---|
| Build 9 / Build 10 | True frozen prospective forecasts; original bytes and accuracy records preserved. |
| V3B1 Preseason Replay | Retrospective research reconstruction with frozen preseason inputs and zero completed games. |
| V3B1 Launch Snapshot / Subsequent V3 Checkpoints | The V3B1 launch snapshot remains immutable. V3B2 and future V3 checkpoints are saved as their own immutable records and evaluated only against eligible results after each record was published. |