Simulate a faceoff
Pick two takers, click a dot anywhere on the ice, or load a real draw from this season, then drop the puck. The probability comes from the same model behind the faceoff posts — fit on 1.2 million NHL draws — and every piece of it is shown below the number.
The draw
The verdict
of the time, the taker wins this draw against the opponent.
Each bar is one term of the fitted model, in points of win probability, from the taker’s side — and each row names the condition it is pricing on this draw. Orange helps the taker; white helps the opponent. These are the shipped coefficients, and this page checks its own arithmetic against the model on every real draw it carries.
The long run
These are faceoff wins, not wins in the standings: if this taker took all 1,467 draws of a first-line season against this opponent, that is how many more of them he would win than a coin flip.
Only the skill gap is carried into the season number. The rest of the verdict — strength, score, home ice, fatigue — describes this draw rather than a season of them: a power play belongs to the moment, not to the centre, so extrapolating it would credit him for the situation he happened to be in. What those extra wins are worth in goals depends on where and when the draws happen, which is the next section’s question.
Manual mode holds the clock at mid-period and the deeper fatigue clocks (prior TOI, rest) even; a loaded real draw carries its actual values for all of them. The single-draw outcome is a weighted coin flip at the model’s probability — except on a loaded real draw, where it goes the way it actually went.
How the model works
One model, 1.2 million draws, 16 seasons. It is a paired-comparison fit: every draw is taker vs taker, so a player’s rating is adjusted for who he actually faced — beating Bergeron counts for more than beating a winger taking his first draw. Below, the situation effects it learned, with the draw you set above marked on each one; the tables come from the published analysis.
The dots are not equal — and it’s the side, not the zone
Home win rate at each of the nine dots, 1.2M draws. Everyone asks about offensive vs defensive zone; the answer is it barely matters (.516 vs .515). What matters is which side of the ice: left dots favor the home taker by five points over right dots, because handedness plays against the dot — see the next panel.
Handedness is about the dot, not the opponent
Win rate by the taker’s hand and the dot. A lefty on the left dot wins .549; the same lefty on the right dot wins .462 — a nine-point swing for the same player. Meanwhile lefty-vs-righty as a matchup is flat: .512–.515 across all four pairings. The folk claim is about the wrong variable.
Strength state
Power-play units win draws — .550 at 5v4, .632 at 5v3. This has to sit in the model’s zero point, or every penalty-kill centre would be scored as bad at faceoffs when he is merely outnumbered.
Fatigue — the icing experiment
Win rate by how much longer the home taker has been on the ice than his opponent. On ordinary draws the effect is almost invisible — tired players change. After an icing the tired team is trapped, and the effect appears: a seven-point monotone slide across the bins. The rule change is a natural experiment hiding in the data. (The fitted fatigue term the sim uses is the small all-draws version; the icing column is the descriptive read that proves the mechanism is real.)
And the takers themselves
The best and worst of the 291 takers the model currently rates (at least 100 effective draws at the as-of date), shown as win probability against a league-average taker on neutral ground. Faceoff skill is the most repeatable thing measured on this site (split-half r = 0.71 — higher than shooting or blocking), and the all-time leader is exactly who a Flyers fan would hope.
What is a won draw worth?
Winning the draw buys possession, and possession decays fast. I measured the xG swing after 318,734 draws, comparing won vs lost between matched takers. The window is the honest researcher’s knob: how long after the draw do you keep crediting the win? Drag it, toggle the game state, and watch the numbers reprice — the combined all-situations read at the 15-second headline is the 0.0144 every other number on this page uses.
The attribution window
Both teams keep generating chances after every draw — winning shifts the split. The dashed line is the value of winning: the full swing between the two futures, the xG you gain and the xG the other side no longer gets, which is why it runs about twice the gap between the solid lines. It saturates past 30 s — the possession a draw buys is spent within half a minute, and fifteen seconds captures 86% of everything there is, which is why it’s the headline and not an arbitrary choice. The placebo sits at −0.0008: run the same 15 s window backwards from the draw and the method finds essentially nothing, which is what it must find if the comparison is fair. The full method.
| game state | your xG if won | if lost | the swing (value) |
|---|
How the value is made, at the window chosen above: both teams generate xG after every draw; the value of winning is the full swing between the two futures — the xG you gain and the xG the other side no longer gets — which is why it is roughly twice the win-minus-lose gap in the first two columns.
Where the value lives
Not all faceoffs are made equal. The 0.0144 above is the average over every draw in the game — break it apart and the price of winning moves by a factor of five depending on where the puck drops and how many skaters are on the ice:
One ruler, two splits, all measured at the 15-second headline window (measured there only, so this chart deliberately ignores the slider). By zone: offensive and defensive draws land on top of each other — at one end you win the chance, at the other you deny it, and in xG terms that is one transaction — while the neutral zone sits far left at a third of the value. By strength: a 5v5 draw is worth 0.0112 and a special-teams draw 0.0266 — about 2.4× — with the all-situations 0.0144 sitting between them as the usage-weighted mix. Practical upshot: you cannot shelter a bad faceoff man by zone, only by taker — which is exactly the Zegras question; what you can do is care most about who takes the special-teams draws.
Your draw, priced
So what is a faceoff man worth?
| skill vs avg | extra wins | goals/season | |
|---|---|---|---|
| a good centre (+1 sd) | +5.3 pts | +78 | +1.12 |
| an elite taker (95th percentile) | +11.0 pts | +161 | +2.32 |
| Claude Giroux (the best) | +16.7 pts | +244 | +3.51 |
Every row is the same thing: how far that taker is above a league-average one, in points of win probability — not one player minus another. On a first-line workload of 1,467 draws, priced at the all-situations 0.0144 per won draw — and both of those are typical-usage mixes: the workload is the draws a first-line centre actually takes across 5v5, the power play, the penalty kill and every zone, and the price is the average over that same mix, so the goals column already reflects real deployment rather than any single strength. Pick a taker in the sim above and his row appears here. Read as a gap instead, the 291 qualified takers span 16.9 points from the 5th percentile to the 95th — 3.56 goals, or 0.6 of a win, over a first-line season — and 28.6 points from the very best to the very worst, though nobody hands the worst taker 1,467 draws. That is the whole honest story: faceoff skill is the most repeatable skill in the model, and the entire deployed spread of it still comes in under a win a season.
Ratings: faceoff__756c1483, fit on 1,205,475 draws through
2026-06-14. Draw value measured on 318,734 draws, 2021–22 to 2024–25. Headshots
from the NHL’s public asset host. Scope: faceoff skill only — the probability this
player wins a draw, adjusted for opponent, dot, handedness, strength, size and fatigue. It
says nothing about what happens after the puck hits the ice; that is the value layer’s
question, priced above.