How the faceoff model works
Faceoff percentage mixes the taker's skill with who he faced, where, and from which side. This page explains how we separate them, how well it predicts draws it has never seen, and how it becomes a projection of faceoffs won.
The model
A logistic Bradley–Terry model, one row per draw, from the home taker's point of view:
logit P(home wins) = θ_home − θ_away + γ_home·S_home − γ_away·S_away + b_home + b_zone + b_centre + b_side·(S_home − S_away) + b_pp·sign(skater difference)
- θ is each taker's skill. It is shrunk toward the league (σ ≈ 0.2 logit) and harder for part-time takers, whose own mean is fitted separately because wingers who take draws after a wave-out are a selected group.
- S is +1 when the taker is on his strong side (a left shot at a dot on his left, facing the net he attacks), −1 on his weak side, 0 at centre ice. γ is a small player-specific strong-side bonus.
- We take the dot from the faceoff's coordinates. Dots sit at x = ±69 and ±20 ft, y = ±22 ft, and we check the orientation against handedness: the strong-side effect would flip sign if it were wrong.
- The fit covers the current season and the three before it, weighted 1, 0.75, 0.5 and 0.35. Early in a season the previous season keeps full weight until the new one has enough draws.
- The rating on our pages is a taker's expected win % against the draw-weighted average regular centre at a neutral dot. Its interval is the 90% posterior interval of θ.
| Home ice | +0.6 pts |
|---|---|
| Own strong side (each taker) | +3.4 pts |
| Extra skater (power play) | +4.8 pts |
| Offensive zone (home) | −1.2 pts |
| Defensive zone (home) | +1.2 pts |
| Centre ice | −0.4 pts |
Does it predict?
We chose the model's settings (the shrinkage strengths) on 2023-24 → 2024-25 only. Then we trained on 2023-24 and 2024-25 and scored every draw of 2025-26 (73,496 draws), a season nothing was tuned on, comparing the probability each method gave the actual winner.
| Method | Log-loss |
|---|---|
| Coin flip | 0.69315 |
| Raw head-to-head records (only draws whose pair had 10+ earlier meetings, from the research sample) | 0.718 |
| Each taker's raw FO%, combined | 0.68585 |
| This model | 0.67974 |
Stored protocol: lambdas tuned on 2023-24 -> 2024-25 (fo-bt-1.1.0); 2025-26 untouched.
Faceoffs stay close to coin flips: even a strong edge moves one draw only a few points. The gains add up over a season and over a game's 20 draws, which is where projections are useful.
What is skill and what is noise
| Measure | Split-half r | Sample | Verdict |
|---|---|---|---|
| FO% (regular takers)Shrinkage point about 130 draws. True-talent spread about 4.8 points. | 0.76 | 417 draws per half | Stable |
| Pair-specific edge (beyond skill and side)SD about 2.1 points; shown shrunk by n/(n + 500). | — | Pairs with 20+ draws, 3 seasons | Noise |
| Arena (scorer) effect on home FO%Spread 0.9–1.2 points, the size of binomial noise. We do not arena-adjust faceoffs. | −0.01 / −0.13 year to year | 32 arenas, 3 seasons | Noise |
From ratings to projected faceoffs won
- Game draws. Both teams take part in every draw, so a game has one draw count: the average of the two teams' recent draws per game, shrunk toward the league.
- Who takes them. Each taker's draws per game over his team's last ten games (newer games count more), normalized so the team adds up to the game total. Part-time takers are discounted by how often they took a draw at all. A taker must have drawn in one of his team's last two games.
- Win rate. The model's probability against each opposing taker, weighted by that taker's share, over his own mix of strong-side, weak-side and centre-ice dots. Same-handed opponents are on the other side of the dot; opposite-handed opponents on the same side.
- Range. Draw counts vary less than a Poisson count (SD ≈ 0.2 × mean), and wins are binomial on top. We report the 10th to 90th percentile.
Backtest: we replayed 2025-26 game by game (1,259 games, 21,132 taker-games), projecting each game only from earlier games and a model trained on earlier seasons.
- Mean absolute error of faceoffs won: 1.57 per taker-game, against 1.56 using the same draws with each taker's raw FO% to date.
- Win-rate log-loss per draw: 0.6881 (model) vs 0.6888 (raw FO% to date) vs 0.6932 (coin flip).
- Draws: mean absolute error 2.50 per taker-game.
- 80% of actual results fell inside the 10th–90th range (target 80%). The range's width was set on the 2024-25 backtest, not this one.
- Opponent swing: for a high-volume centre, the expected win rate ran from 51% (toughest tenth of matchups) to 58% (easiest), about 1.2 faceoffs won per game between them.
Read plainly: the model predicts win rates better than percentages to date, but on faceoffs won per game it is no better than that simpler projection. Most of the per-game error comes from the number of draws (penalties, icings, a centre moved to the wing or scratched), not from the win rate. The model's edge shows in head-to-head probabilities and ratings, not in a single game's count. We have not separately tested weekly totals.
Limitations
- The feed records only the taker who took the draw. Violations and wave-outs, tie-ups, and the quality of a win (clean to a defenceman vs a scrum) are not captured.
- Projections mix opposing takers by their share of draws, not by line matching. A coach who matches lines can tilt who faces whom.
- A lineup change (injury, scratch, trade) reaches the projection after the next game the team plays. In the backtest about 5% of projected takers (3+ expected draws) took no draw that night.
- Score state and fatigue (shift length) are not in the model yet.
FAQ
Does handedness matter on faceoffs?
Yes, more than anything else we measured except the takers' own skill. A taker on his strong side (a left shot on his left dot) beats a taker on his weak side 56.7% of the time. Same-handed centres swing about 14 points between the two dots of a zone; opposite-handed centres are always both strong or both weak, so the dot does not matter for them.
Is a faceoff head-to-head record meaningful?
Mostly not. After controlling for each taker's skill, home ice, zone, dot side and strength, pair-specific edges have a standard deviation of only about 2.1 points. Predicting from raw head-to-head records did worse out of sample (log-loss 0.718) than a coin flip (0.69315). We show the raw record next to the model and shrink it heavily toward the model.
Is there a home-ice advantage on faceoffs?
A small one: home takers won 50.44% of draws in 2025-26, and after adjusting for who took them the edge is about a point or less (it depends on the seasons in the fit; the methodology page shows the current value). Rule 76.4 (the visitor puts his stick down first at centre ice) shows no measurable edge, and arena scorers show no bias on faceoffs.
Are faceoffs a skill?
Yes. Split-half reliability of FO% is r = 0.76 at about 417 draws per half, which puts the shrinkage point near 130 draws. Observed spread between regulars is about 5.4 points, of which about 4.8 is talent.
Model fo-bt-1.1.0, trained on 2026-27, 2025-26, 2024-25, 2023-24. Faceoff leaders →