The same rate can carry different evidence
Imagine two shooters facing chances rated at 10% each. These are hypothetical examples.
| Record | Goals | Expected goals | Scoring rate |
|---|---|---|---|
| 10 attempts | 3 | 1 | 30% |
| 100 attempts | 30 | 10 | 30% |
Both scored at three times the expected rate. The second record gives us much more evidence that the difference deserves attention. Three goals in ten attempts can swing sharply with one extra save or a shot off the post.
The model therefore puts more weight on the player’s own result when the evidence is stronger, and pulls more toward average when it is weaker. It moderates poor results as well as good ones.
What “empirical Bayes” adds
The method uses the wider player population to estimate how much shooting effects vary. It combines that reference with the information in the individual player’s attempts.
If players appear to differ substantially, the calculation allows more separation. If much of the observed spread can be explained by ordinary variation, it pulls individual estimates closer together.
The number of attempts matters, but so do their scoring probabilities. There is no universal rule that every player keeps half his goals above expected, or that a particular shot count makes an estimate certain.
How it becomes a finishing number
We start with the gap between goals and expected goals, estimate an effect on scoring odds, and shrink that effect toward zero. Zero here means no extra scoring effect beyond the chance baseline.
We then apply the adjusted effect to each of the player’s chances and add up how much their scoring probabilities change. That produces estimated finishing goals, which are converted into wins.
Working with odds lets the adjustment keep probabilities between zero and one. It also means we cannot reproduce the result by multiplying the raw goal gap by one fixed percentage. The proportion of the raw gap that remains is not a probability that the player’s skill is “real.”
The adjustment is made for each player-season. It does not separately reward a player for an apparent run of consistency across several years, and it does not turn the result into a next-season forecast.
Why we use it instead of crediting the whole raw gap
Goals above expected is useful: it tells us how far scoring finished above or below the model’s expectation. But that gap combines the shooter’s contribution with goaltending, ordinary variation and details the chance model did not capture.
Crediting the entire gap to the shooter would treat all of those influences as his contribution. Shrinkage reduces that leap. It helps us avoid letting a short scoring burst—or a run of misses—dominate the assessment.
There is a cost: it can understate a truly exceptional performance. It does not identify exactly which goals were luck, and we have not established one ideal amount of adjustment for every player and purpose.
That is why the raw result and adjusted finishing estimate answer different questions. We keep the raw gap available as a description of the scoring result; we use the adjusted estimate when assigning finishing credit within WAR. The goals still happened. The question is how confidently we can attribute the entire difference to the shooter.