How HGB
Talent Works

Expected goals (xG) is the engine under almost everything here. Every unblocked shot — on net or missed — gets a probability of becoming a goal, based on where it was taken, the angle, the shot type, whether it followed a rebound and how fast, whether it came off the rush, and what happened the moment before. Add those probabilities up and you get chance quality that doesn't care whether the puck actually went in.

HGB runs four separate xG models, not one: even strength, power play, penalty kill, and empty net. A power-play shot and a 5v5 shot are different problems, with different shot locations and different goalie behavior. Push them through a single model and it over-rates power-play chances and flattens the rest. Splitting them tightens the calibration: predicted goals now land within a few percent of actual goals in every situation, and within about a percent on special teams and empty net.

The models are gradient-boosted, trained on every unblocked shot since 2022 — the seasons after the NHL changed its shot tracking, so the location data is consistent. This is the same shot basis the public models use: unblocked attempts, missed shots included. xG feeds the even-strength and power-play numbers below and the goalie save model. Better xG, better everything downstream.

RAPM (Regularized Adjusted Plus-Minus) isolates individual player impact by controlling for teammates, opponents, zone starts, score state, and PP expiry. Every shift creates a set of equations where each player on the ice gets credit or blame for what happened, and ridge regression finds the player-level estimates that best explain the full dataset — while shrinking noisy estimates toward zero.

A player with a small sample gets pulled toward league average. A player with thousands of 5v5 minutes gets a more confident estimate. The model uses a two-perspective design: each shift generates one home-offensive row and one away-offensive row, keeping offensive and defensive signals separated so that a defenseman who suppresses shots doesn't accidentally look like an offensive contributor.

WAR measures what a player produced; Talent estimates how good he is. One major difference is deployment. WAR credits every minute a player skated, so heavy power-play time gives a player more opportunity to accumulate WAR. Talent normalizes that: each situation (even strength, power play, penalty kill) is weighted by a fixed position-level share of ice time rather than the player's own minutes. Extra power-play minutes don't mechanically increase Talent; they provide more evidence about the player's per-60 performance.

Talent leans on history. Each component blends the current season with up to three prior seasons, weighted toward the recent ones (roughly 30 / 16 / 8 across the last three years) and adjusted for age. How far the current season pulls the estimate depends on sample size — the thresholds below mark where the current season and the prior carry equal weight.

Losing a role doesn't erase ability. A player dropped from the power play keeps his established power-play prior instead of collapsing to zero; the missing minutes lower his confidence rather than his score. Players with thin history (under 1,000 even-strength minutes across the last three seasons, or a player with no usable prior history) are flagged limited sample on the card.