FHI is Full Press Hockey’s season-level estimate of an NHL skater’s overall impact, presented on a 0–100 scale. It combines adjusted offense and defense built around our own expected-goals model, FPxG, with special-teams contributions and reliability adjustments.

Use it to explore how a player influences the quality of chances created and allowed. It is a model estimate of performance in the selected season—not a complete scouting report, a points ranking, or a guarantee of future results.

• FHI — Full Press Hockey Index: the overall player impact index.
• FHOI — Full Press Hockey Offensive Index: the offensive impact index.
• FHDI — Full Press Hockey Defensive Index: the defensive impact index. Higher means better estimated chance prevention.
• FHV — Full Press Hockey Value: cumulative modeled expected-goal value, which incorporates playing time.

The three indices describe impact rates relative to a reference group. FHV describes how much modeled value accumulated over the player’s actual modeled minutes.

The scale is 50 + 15 × z, limited to 0–100, where z is the raw impact rate’s distance from its reference-group mean in standard deviations. A 50 sits at that reference center; 65 is one standard deviation above it; 80 is two above it.

This is not a percentile, school grade, or probability. An 80 does not mean “better than exactly 80% of players,” and it does not mean twice the impact of a 40. In the provisional 2026–27 season, the reference is frozen from 2025–26, so the current league average does not have to equal 50.

No. The model adds the underlying offensive and defensive rates, then normalizes that total separately to produce FHI. FHOI and FHDI each have their own reference mean and spread. Averaging the two displayed component scores will not reconstruct FHI.

Both use 0–100, but forwards and defensemen are normalized separately. A defenseman’s 75 describes performance relative to the defenseman reference distribution; a forward’s 75 describes performance relative to forwards. Equal displayed scores therefore do not necessarily mean equal raw expected-goal impact.

Use the position filter for the clearest ranking comparison. When comparing roles across positions, examine the underlying rates, usage, and FHV as well.

FHI, FHOI, and FHDI are normalized rate indices. FHV is cumulative: FHV = total impact rate × modeled minutes ÷ 60.

Illustration: a player producing +0.60 modeled expected goals per 60 over 600 modeled minutes has +6.0 FHV. Another producing +0.40 over 1,200 minutes has +8.0 FHV. The first has the higher raw rate; the second has accumulated more value. These are illustrative numbers, not player results.

No. Index values are capped at 100 and floored at 0 for presentation. A 100 can conceal differences between exceptionally high raw rates; it is not a perfect-player claim. A negative raw rate or FHV means below the model’s baseline over that sample—not negative actual goals, zero NHL ability, or a verdict on a player’s contract.

The positional center used to display an index and the model baseline used for raw value are different concepts. An FHI of 50 does not necessarily imply exactly zero FHV.

The public player FHI is a skater model. It does not provide goalie FHI ratings. Chance prevention by skaters is different from a goalie’s save performance.

FHI shares the Full Press index presentation approach with football’s FQI, but the formula is hockey-specific. Do not apply FQI component percentages, qualification rules, or score interpretations directly to FHI.

The pipeline uses NHL-origin play-by-play, box scores, and shift charts. These supply shot context, player identities, game state, and on-ice personnel. Full Press calculates FPxG and the impact indices from those inputs; it does not relabel another analytics site’s expected-goals numbers.

Source corrections can change historical inputs and subsequent ratings. A finished game is eligible only when the necessary data can be processed successfully.

FPxG is Full Press expected goals: the estimated scoring probability of an eligible shot given the recorded circumstances. A 0.10 shot contributes 0.10 expected goals. Ten such chances sum to 1.0 expected goal, but that does not guarantee one actual goal.

Individual FPxG counts the player’s own chances. On-ice FPxG counts chances for the player’s team or opposition while the player is on the ice. Neither is the same thing as the adjusted FHI impact coefficient.

The model uses shot distance and angle, shot type, strength state, the preceding event, elapsed time, movement between recorded events, lateral movement, rebound and rush proxies, pre-shot score, period, and nonlinear combinations of those features. It uses gradient-boosted trees to estimate scoring probabilities.

Shooter identity, goalie identity, and team identity are excluded from the shot-probability model. A star’s name does not automatically make the same recorded chance more valuable. Recorded-event proxies are not equivalent to complete puck-and-player tracking.

Standard FPxG covers eligible unblocked, non-shootout, non-empty-net attempts. Blocked-shot coordinates can describe the block location rather than the original shot, so blocked attempts are not assigned standard FPxG. Attempts without usable coordinates are held out rather than silently assigned an invented location.

Empty-net events can remain in raw game accounting without entering standard FPxG. The adjusted player backbone uses exact five-on-five stints with both goalies present and five identified skaters per side. There is no dedicated three-on-three or four-on-four player impact model in this version; broader individual shooting refinements can still draw on eligible shots outside exact five-on-five.

The five-on-five backbone uses regularized adjusted plus-minus (RAPM). Play is split into stints with unchanged personnel, with two observations per stint: one for each team’s offense. The model estimates FPxG created per 60 while accounting together for the five attacking skaters, five defending skaters, home ice, and score differential clipped between −2 and +2. Stints are weighted by duration.

Ridge regularization reduces extreme estimates when samples are small or players share many minutes. It helps separate contributions; it cannot perfectly identify who caused every chance.

The current five-on-five model explicitly accounts for skater identities, home ice, and score state. It does not include separate zone-start, back-to-back, injury, age, or coaching variables. Some effects can appear indirectly in observed results and shared personnel, but that is not the same as explicitly controlling for them.

Deployment and systems remain part of the interpretation. Avoid describing FHI as a fully context-free measure of talent.

Power-play offense starts with the player’s on-ice PP FPxG-for per 60 minus the league PP rate. Penalty-kill defense starts with the league PK FPxG-against rate minus the player’s rate, so better prevention is positive. Each is shrunk toward league average using minutes ÷ (minutes + 120).

The model weights five-on-five, PP, and PK contributions by actual modeled minutes. Special teams use league-relative on-ice rates, not the same player-isolating RAPM fit used at five-on-five. PP defense and PK offense are not separate impact components in this version.

The offensive refinement adds two conservative components. First, it estimates individual FPxG creation beyond what is already explained by the player’s adjusted offense and power-play share, separately for forwards and defensemen. That residual is shrunk using a 750-minute prior and blended at 0.45.

Second, goals above FPxG per 60 are shrunk using a much larger 3,000-minute prior and blended at 0.35. This gives finishing some credit without treating a short hot streak as established skill. The creation residual helps reduce double counting of shooting already reflected in on-ice offense.

The validation work found defensive estimates less stable across split samples than offensive estimates. FHDI therefore applies an additional reliability adjustment based on five-on-five plus PK minutes, with a 1,100-minute prior and a 0.65 partial blend toward zero.

This makes small-sample defensive extremes more conservative. It does not mean defense is unimportant, and it does not make the remaining estimate certain.

Yes. All rates below are per 60 minutes. Let E, P, and K be modeled exact-five-on-five, power-play, and penalty-kill minutes; M = E + P + K. Let T be total player minutes used for the individual shooting refinements.

The creation residual comes from a positional regression on base offense and PP deployment share. Offense, defense, and total each receive their own positional normalization. The 0.45, 0.35, and 0.65 constants are blending factors, not percentages that sum to 100%. Playing-time weights depend on the player’s actual usage. Zero-exposure components do not create value.

As of this FAQ’s October 7, 2026 review, the public product includes 2024–25, 2025–26, and provisional 2026–27. The season selector on the rankings page is the authority for currently available releases.

The hockey season view combines NHL regular-season and postseason games. It is not a separate regular-season/playoff scope switch like some football pages. Preseason, All-Star, Olympic, and other international games are excluded.

Current-season inputs and player coefficients are estimated from the new season, while ridge penalties, the creation-regression calibration, and index reference means and spreads are frozen from the reviewed 2025–26 reference. This prevents an extremely small early-season pool from defining a completely new scale every day.

Prior-season player FHI scores are not carried forward as the new season’s player ratings. The season remains provisional under this workflow until a separately reviewed recalibration or release changes that status.

200 exact-five-on-five minutes define the research qualification pool and the historical positional normalization reference. 500 exact-five-on-five minutes are the headline public-ranking threshold. These are five-on-five minutes, not total ice time or games played.

The model can calculate lower-sample results, and those players remain discoverable through All Skaters/search. Qualification filters reduce noise; crossing a threshold does not suddenly make a rating precise or prove a player improved.

A calculated score and qualification for a headline rank are separate things. A player below the required threshold can have a displayed index while a qualified-rank field remains a dash.

The current-season All Skaters presentation does not repeat a Small Sample badge on every player. That cleaner display does not waive the qualification thresholds or the provisional status. Check minutes and the selected pool; historical views may still use sample labels.

FHI is a scheduled season-model refresh, not an in-game scoreboard. The current pipeline is scheduled twice daily and processes official/final games through the previous calendar day in America/New_York time. The published cutoff and update information on the page are the authority for what a snapshot includes.

If an upstream feed or model run fails validation, the previously published snapshot can remain visible. A final score appearing elsewhere on the site does not prove that game has already entered FHI.

Adjusted impact is estimated jointly across the season. New games can change estimates for teammates and opponents and therefore the separation of earlier shared minutes. Source corrections, completed shifts, and recalculation can also change a player’s result.

In historical season-relative normalization, the reference distribution can change after a rebuild. In provisional 2026–27, the frozen reference scale removes that particular source of daily movement, but jointly estimated coefficients and data corrections can still move ratings.

The current-season pipeline aggregates performance by player ID across the season. The displayed team is the most recent team in that player’s processed game record. A new team badge therefore does not mean all of the displayed season’s impact occurred with that team.

Do not interpret a team filter as a guaranteed before-and-after trade split. For that question, inspect the game log and use a dataset explicitly separated into team stints.

Compare with care. Historical indices describe position-relative standing in their season’s reference distribution; current 2026–27 uses the frozen 2025–26 reference. League scoring environments, coverage, model versions, and deployment can differ.

The same displayed FHI across seasons does not automatically mean identical raw impact or talent. Match the formula version and scope, then consider raw rates and minutes. The 2024–25 FPxG results are retrospective applications of a model trained on later 2025–26 data, not forecasts made before those 2024–25 games.

Points describe recorded scoring production. FHI also evaluates adjusted chance creation, chance prevention, usage, and the reliability of finishing above expected. A player can collect many points with excellent linemates or heavy PP opportunity while another contributes strongly at five-on-five or defensively.

A disagreement is a reason to inspect components and game context. It is not proof the points are meaningless or that the model has discovered the unquestionably better player.

Yes. FHOI and FHDI intentionally keep those dimensions visible. An excellent FHI can be driven primarily by one side of play. Read the component scores before calling someone a complete two-way player.

A defensive specialist with limited PP time and a high-usage offensive player can reach similar overall values through very different contributions.

No. FHDI reflects estimated prevention of expected goals, with positive defense meaning fewer dangerous chances allowed relative to the modeled baseline. It is not a save-percentage award or a count of blocks, hits, or takeaways.

A player can defend well by preventing attacks before a shot occurs. Another can accumulate blocks because the team spends extended time defending. Those observations require context beyond the raw count.

There is no direct points-style bonus for assists, hits, blocks, takeaways, or faceoff wins. Playmaking and off-puck work can affect modeled on-ice chance creation and prevention, but the model does not assign a separate tracked value to every pass, screen, entry, or recovery.

Penalties drawn and taken do not have standalone value components in this version. Special-teams outcomes and usage are included, but that does not fully credit the individual who created or conceded the manpower advantage.

No fixed bonus is awarded for having a role. Contributions depend on rates relative to the applicable baseline, sample shrinkage, and actual minutes. However, opportunity and teammates still matter, especially because the PP/PK inputs are on-ice league-relative measures.

A player with no PP or PK exposure receives no observed contribution from that component. Lack of opportunity should not be read as proof the player could not succeed in that role.

Do not treat a narrow gap as a demonstrated talent difference. Small samples, shared linemates, incomplete context, and model choices introduce uncertainty. The public index does not currently provide a confidence interval establishing that every adjacent ranking is statistically distinguishable.

Start with exact-five-on-five minutes, then check component balance, usage, and longer-term evidence. Early-season leaders are useful observations to follow, not definitive season awards.

Rounding and the 0–100 caps can hide differences. The ranking uses the selected index, then the corresponding uncapped raw impact rate, then FHV, with a stable player identifier as the final player tiebreaker. Team rankings use an analogous process with the team abbreviation as the final fallback.

A visible tie therefore need not be a tie in the underlying model.

FHI can add context, but those are different questions. Fantasy scoring may reward goals, assists, blocks, or saves in combinations FHI does not target. Contract evaluation needs age, health, cost, role, and future expectations. Awards can value availability and cumulative contribution as well as rates.

FHI is not a calibrated betting probability or a projection of a player’s next game. Use FHV for accumulated model value and combine the model with appropriate evidence for the decision.

Corsi counts shot attempts, including blocked attempts; Fenwick counts unblocked attempts. Their percentage forms describe a team’s share of those attempts while a player is on the ice. Evolving-Hockey’s definitions.

FHI uses shot-quality probabilities instead of giving every attempt equal value, applies a five-on-five adjusted-impact model, adds special teams and individual refinements, and normalizes the result. It is not a renamed CF% or FF%. The simpler measures remain useful for understanding territorial results.

On-ice xG% is generally xGF ÷ (xGF + xGA): the share of expected goals belonging to a player’s team during the selected minutes. It describes those shared on-ice outcomes. On-ice metrics and percentages.

FHI is an adjusted player impact estimate, not a share. A 60 FHI does not mean a 60% xG share. When using Natural Stat Trick or any other table alongside FHI, align the season, strength state, score/venue adjustments, minimum minutes, and xG provider before comparing. Different filters alone can produce very different leaderboards.

Both are expected-goals models, but they are independently built. MoneyPuck documents its own shot features and also offers separate flurry-adjusted and shooting-talent-adjusted variants. Those variants should not be treated as interchangeable with ordinary xG. MoneyPuck methodology.

FPxG uses Full Press’s own training, event eligibility, and feature processing. FHI then builds adjusted impact on top of FPxG. Differences in training seasons, shot handling, and model calibration can change the result. FHI’s separate finishing refinement does not turn FPxG itself into MoneyPuck’s talent-adjusted model.

FHI uses the RAPM family of methods for its five-on-five backbone. RAPM is the statistical approach; FHI is the complete Full Press pipeline and presentation, including its own FPxG inputs, defined controls, special-teams treatment, shooting refinements, reliability rules, and index scale.

It is not identical to an Evolving-Hockey RAPM chart. That provider documents different model specifications, including zone-start and rest variables absent from this FHI version. Evolving-Hockey RAPM methodology.

GAR and WAR estimate value above a defined replacement baseline in goals or wins. Evolving-Hockey’s framework separates contributions such as even-strength offense/defense, special teams, and penalties; expected-goals variants use their own specifications. Evolving-Hockey GAR/WAR documentation.

FHI is a position-normalized impact-rate index. FHV adds playing time, but it is measured against the FHI model baseline and has no replacement-level adjustment or goals-to-wins conversion. FHV is not GAR or WAR. You cannot convert a 75 FHI into a specific WAR total by a fixed multiplier.

The JFresh player-card documentation at HockeyStats.com describes projected WAR metrics using weighted averages of the three most recent seasons, displayed as percentile rankings. Player-card methodology.

FHI describes the selected season’s observed sample with its own adjustments and reliability rules. Its 0–100 scale is a transformed z-score, not a percentile. The provisional 2026–27 reference uses prior-season calibration, but it does not blend three seasons of each player’s FHI into a future-performance forecast. Compare the time horizon and scale before interpreting a disagreement.

A single-game card addresses what happened in one game under that card’s formula and scope. HockeyStatCards, for example, presents game-score breakdowns and identifies its model attribution.

FHI is estimated over the selected season. Its public game logs show the underlying FPxG process and usage, not a separately calibrated single-game FHI. Do not average a set of game-card ratings to recreate season FHI, and do not assume every product using “Game Score” or “GSVA” shares one formula or forecasting horizon.

Point Shares estimates a player’s contribution to team standings points. Hockey-Reference’s Point Shares explanation.

FHI instead summarizes adjusted expected-goal impact as a normalized rate; FHV accumulates modeled expected-goal value. Neither number is an allocation of actual standings points. A win, overtime loss, or team’s place in the standings is not directly translated into a player FHI bonus.

Tracked microstats can describe actions such as entries, exits, passing, and rush defense. HockeyStats.com’s microstat cards document measures built from All Three Zones tracking. Microstat-card methodology.

Those data help explain how a player produces an impact. FHI does not directly ingest a complete tracked pass, entry, or off-puck dataset. Use video and microstats alongside FHI when an outcome-based rating leaves the mechanism unclear.

These approaches observe different parts of performance. FHI focuses on expected-goal process and adjusted contributions, with conservative treatment of finishing. Box-score totals reflect recorded outcomes; video reveals decisions and context that public event data can miss.

Investigate the mismatch: sample size, teammates, competition, special teams, recent role changes, chance quality, finishing, and unmeasured defensive work. The model can be wrong, and a memorable highlight can also distort a human impression. Neither should automatically override the other.

No universal superiority claim follows from this methodology. A fair comparison needs the same target, game sample, evaluation period, and genuinely held-out data. Predicting future goals, describing past value, isolating defense, and explaining a player’s style are different tasks.

FHI’s distinctive choices are its own FPxG foundation, exact-five-on-five adjusted backbone, specified shooting refinements and shrinkage, and transparent component presentation. Those choices provide another informed perspective; they do not make other established models obsolete.

FHTI is the Full Press Hockey Team Index, the team’s normalized overall modeled impact rate. FHSTI is the Full Press Hockey Special Teams Index, the normalized special-teams value rate.

The team rankings combine five-on-five process and special-teams performance. Team offense and defense components help explain the overall result.

No. The team model is fitted separately using team-game FPxG rates, team offense, opposing-team defense, and home ice to account for schedule context. It then combines those estimates with special-teams contributions using actual minutes.

Player indices use separate positional normalization and individual refinements. Adding or averaging those displayed indices does not reconstruct a team index.

Both use league-relative FPxG rates and observed usage, but the team formula does not apply the player component’s 120-minute PP/PK shrinkage. The player creation, finishing, and defensive reliability refinements introduced in formula v0.5.1 also do not alter the team backbone.

This is why a team result and a collection of player components can differ even when they come from the same games.

FHTI primarily evaluates modeled chance creation and prevention. Expected-goal defense is not the same as actual saves, and the model is not a standalone goalie evaluation. A team can suppress chances well yet allow more actual goals, or win behind excellent goaltending despite weaker chance results.

FHTI is neither the standings table nor a calibrated playoff or Stanley Cup probability. Full Press editorial power rankings can incorporate a broader assessment and need not match FHTI.

The logs expose game-level FPxG process and usage. The season index comes from the team model fitted across the selected season and then normalized. The public product does not label an uncalibrated game result as an official single-game FHTI.

A single strong game can be informative without replacing the season’s larger body of evidence.

Select the same season, position group, and qualification pool. Compare exact-five-on-five minutes first, then FHI, FHOI, FHDI, and FHV. Check PP/PK deployment and the individual/game-log context. Distinguish “higher rate” from “more accumulated value.”

Finish with video, role, health, and other independent measures. Keep the size of the gap and the uncertainty in view, especially early in the season.

Check the season, spelling, team and position filters, and qualification selection. Try All Skaters where available. A player may have no eligible processed games or insufficient usable shift/event data. Goalies are outside the skater leaderboard.

A missing value is not automatically zero performance. The published data cutoff can also explain why a recent debut or new team assignment has not appeared yet.

Per 60 standardizes a count or rate to 60 minutes of the relevant ice time: count × 60 ÷ minutes. Always match the numerator and denominator to the same situation.

FHI qualification uses exact-five-on-five minutes. FHV uses modeled five-on-five + PP + PK minutes. Broad individual season shooting refinements use total player minutes. These denominators are not interchangeable, and a short sample can produce an extreme rate.

No. A log’s shot, FPxG, and usage fields explain game-level process. The official FHI is a season estimate fitted across the whole included sample. A daily refresh is a new season snapshot, not a separate player rating for that day’s game.

Do not subtract two rounded FHI scores to claim an exact amount of value added in one game.

Include the player, selected season, FHI/component or FHV value, qualification pool, relevant minutes, and the page’s data cutoff. Link to the player rankings or team rankings and this methodology page.

For a suspected error, record the affected game/date, filter settings, and expected versus displayed value. That helps distinguish a source-data issue, an update delay, a filter mismatch, and a legitimate model disagreement.

FPxG v1.0.0 was evaluated using a chronological holdout from corrected 2025–26 regular-season and postseason shots, then refitted on the full corrected sample. Player-model review examined split-sample stability, component behavior, shrinkage, score distributions, and qualification thresholds.

Those checks support the chosen implementation; they do not prove perfect player attribution or superior prediction against every competing model. Shot-probability validation is also not the same test as validating an overall player ranking.

Public play-by-play and shift records can contain missing or inaccurate coordinates, event timing, and personnel. Teammates often share large portions of their minutes, making individual separation difficult. Tactical assignments, injuries, screens, passing detail, and off-puck execution are incompletely observed.

Special teams are less individually isolated than the five-on-five backbone. Early samples remain noisy despite shrinkage. Separate positional scales, score caps, retrospective model application, and the absence of displayed uncertainty intervals also constrain interpretation.

No. Adjusted observational estimates are not controlled experiments. Regularization and context variables improve the comparison but do not reproduce every coaching, linemate, or opponent response to a hypothetical lineup change.

FHI should not be read as “remove this player and the team will lose exactly this many goals or wins.” FHV is accumulated modeled value, not a causal replacement experiment.

This page was reviewed on October 7, 2026. It describes the public FHI presentation v1.0, the validated fhi-0.5.1 player formula, FPxG v1.0.0, the corresponding team backbone, and the provisional 2026–27 current-season calibration.

A website/plugin release number can change without changing the scoring formula. Conversely, a future model revision may change historical results after a rebuild. Match the formula, data cutoff, and season rather than assuming every software update changes the mathematics.

FPxGF / FPxGA

Full Press expected goals for / against.

Individual vs. on-ice

The player’s own recorded actions versus team/opponent events during that player’s minutes.

TOI · 5v5 · PP · PK

Time on ice · five-on-five · power play · penalty kill.

RAPM

Regularized adjusted plus-minus: a regression method for estimating player contributions while accounting for modeled context.

Ridge / shrinkage

Techniques that make unstable estimates more conservative. Shrinkage is not a guarantee that an estimate is correct.

Residual

The portion left after accounting for the specified predictors; in the creation refinement, what base offense and PP share do not explain.

z-score

Distance from a reference mean, measured in that reference distribution’s standard deviations.

Rate / cumulative value

Performance standardized for time versus value accumulated over actual modeled time.

Model baseline / replacement level

The reference used inside an impact model versus an explicitly defined replacement-player standard. FHV does not add a replacement-level adjustment.

Use the creators’ documentation, because names, versions, and filters matter:

• Evolving-Hockey: general terms and shot metrics
• Evolving-Hockey: RAPM
• Evolving-Hockey: GAR and WAR
• MoneyPuck: models and expected-goals variants
• HockeyStats.com: JFresh player cards
• HockeyStats.com: microstat cards
• HockeyStatCards: game-score cards
• Hockey-Reference: Point Shares

These links explain those providers’ products. They are independent of FHI and do not imply an endorsement of Full Press or a shared formula.