Blog · Grit & GF, Research
Does grit win games?
We tested Grit & GF the way we said we would. The rating repeats, it isn't a list of stars, and it doesn't predict results beyond expected goals. One part does.
By Chase ·
When we launched Grit & GF we listed three ways it could turn out to be wrong: the parts might not repeat, it might be a talent ranking with extra steps, and none of it might ever show up on the scoreboard. We wrote the test down before running it, with the bar a part had to clear. Here is what came back.
The short version
The rating repeats about as well as points per 60, and its top 25 is mostly not the league's top scorers. But last season's ratings, averaged over tonight's lineup, don't predict results any better than our expected-goals game model already does. One part, chemistry, does, and it held up when we controlled for how talented the lineup was. So Grit & GF is a description of how a player plays. Treat it as a betting signal and you'll be disappointed, with one exception worth chasing.
Does it repeat?
We rated three full seasons with the production method and compared each skater with himself the next year, for the 600-odd players with 200 or more minutes at 5-on-5 in both. Then, to take aging and role changes out of it, we split each season into odd and even games and compared the halves.
Season to season, same player
Correlation between a skater's number one season and the next, 2023-24 to 2025-26, about 610 skaters per pair. Points per 60 and on-ice xGF% are there for scale.
- GF0.96
- Points per 60 (for scale)0.84
- Grit & GF composite0.80
- Physicality0.79
- Chemistry0.55
- On-ice xGF% (for scale)0.52
- Intimidation0.30
- Clutch-0.03
GF repeats almost perfectly, which makes sense: where a player takes his shots from is close to a fixed trait. Physicality and the composite repeat about as well as scoring does. Chemistry sits with on-ice xGF%. Intimidation is shaky. Clutch doesn't repeat at all, from season to season or from one half of a season to the other, so what it measures is luck. The original piece said that was the likely outcome, and it was.
Is it a talent ranking with extra steps?
Partly, in two of the parts. GF correlates 0.63 with points per 60, and chemistry 0.55, because good players make their linemates look good. Physicality has no correlation with scoring at all. The composite lands at 0.46. At the top of the board, only 2 of last season's top 25 by rating were among the league's top 50 scorers per 60. It isn't a list of stars.
Does it predict results?
This was the test that mattered. For every regular-season game of 2024-25 and 2025-26, 2,624 in all, we averaged each team's ratings from the season before over the skaters who dressed, weighted by their minutes. Last season's ratings were finished before this season started, so nothing leaked. The baseline was our walk-forward game model, the same one on the report card, which already knows each team's expected goals, the goalies and home ice. Then we asked whether the grit gap between the two lineups explained goal differential the model missed.
Goals beyond the game model, per SD of the grit gap
Slope of (actual goal difference minus the model's expected goal difference) on the home-minus-away team index, in standard deviations; t-statistic with robust errors; 2,624 games. A part needed |t| of 2.6 or more, the same sign in both seasons, and a better out-of-sample win prediction.
- Chemistryt +3.0
- Intimidationt +1.8
- Composite ratingt +0.6
- Clutcht +0.1
- GFt −0.4
- Physicalityt −0.6
The composite rating comes in at +0.03 goals a game per standard deviation, with a t of 0.6. The test could have caught anything bigger than about 0.14 goals, so this is a real null and not a shrug from a small sample. Physicality is slightly negative, which answers the question the methodology left open, for now: hits and blocks stay worth zero goals. GF adds nothing once expected goals are in the model, which is no surprise, since GF is mostly built from them.
Chemistry is the exception. A one standard deviation chemistry gap between two lineups is worth about 0.15 goals of differential the model didn't see, roughly two to three points of win probability. It was positive in both seasons, and fitting it on 2024-25 improved the win predictions for 2025-26 by 1.7 thousandths of log loss. For scale, the whole game model beats a constant guess by about 13. Intimidation came close and didn't make it: strong in 2025-26, nothing in 2024-25.
Is chemistry just talent again?
That was the obvious objection, and we didn't write it into the plan, so treat this part as an afterthought. We added two lineup talent measures built the same way, last season's points per 60 and on-ice xGF%, and ran chemistry alongside them. Chemistry got stronger, +0.19 goals per SD with a t of 3.1, and points per 60 faded to a t of 1.5. Whatever chemistry measures, it isn't only that the lineup is good. The season split is the soft spot: a t of 3.1 in 2024-25 and 1.0 in 2025-26.
What changes
The site now calls Grit & GF what the data says it is: a rating of how a player plays. It isn't a betting edge and it doesn't touch any price. Hits and blocks keep their zero-goal value. Clutch doesn't repeat, so as of grit-0.5 it counts for nothing in the rating or the goals value; its 8% went to the other four in proportion. It stays on the card so you can see it.
Chemistry gets the next test, and this one runs forward on games nobody has seen yet. Each morning before puck drop a shadow game model adds a chemistry adjustment, fitted on the games above and frozen, using that day's projected lineups. The shadow prices are stored and never shown. After the 2026-27 regular season they get scored against the live model on the same games. It passes if the goal residual on projected-lineup chemistry is above zero with a one-sided t of 1.65 or more and the shadow's moneyline log loss beats the live model's. If it passes, chemistry goes into the prices. If it doesn't, this paragraph gets an update either way. The ratings board stays free.