NHL Game Script Forecast for Sunday, Oct 4

We split every second of every game by score state, using the shift data behind our line tools: who was on the ice, and whether their team was leading, tied or trailing. From that we get each skater's share of their team's ice time in each state, the same split for power-play time, and their share of the team's extra-attacker time (goalie pulled while trailing in the third period or later).

A skater's usage when their team trails, compared with when it leads, is a stable trait. In 2025-26 the odd-numbered games and the even-numbered games agreed at r = 0.73 (583 skaters with 20+ games in each half, 0.85 after the Spearman-Brown correction). The gap runs from about 157 seconds per 60 minutes less when trailing (5th percentile) to 195 seconds more (95th). Trailing teams also get more power-play time: 316.7 seconds per 60 minutes trailing against 250.3 leading.

For tonight's games we estimate each team's win probability. We start from goal differential per game, mixed with last season's rating and regressed toward average, then add home ice and Schedule Edge's rest, travel and backup-goalie effects. We turn that probability into an expected share of the game spent leading, tied and trailing: each 10 points of win probability moves the expected trailing share by about 4.1 points.

Projected ice time is the expected time in each state multiplied by the skater's share in that state. Shares are shrunk toward the skater's all-situations share, and early in the season they're blended with last season's.

What the test showed: we fit on 2024-25 and scored 2025-26 (the coefficients in use were then refit on both seasons, so the test numbers describe the 2024-25-only fit). Pregame strength explains only 0.6% of how long a team actually spends trailing, because hockey scores are random. So the script-adjusted ice-time projection was no more accurate than a neutral one: 130.2 s against 130.2 s mean absolute error over 22,691 skater-games. Even knowing the final score-state mix in advance only improves that to 127.8 s. The win-probability model is only slightly better than home ice alone (log loss 0.6935 against 0.6955) and its test calibration was uneven, so we show win probability rounded to the nearest 5%. We publish the forecast as context, not as an edge: read each skater's trailing and leading usage as scenarios.

See also: garbage-time points · Schedule Edge. Model version gamescript-1.0.0.