Who actually rose in the postseason — measured as a shift in possession-level impact against the player’s own regular-season self, not as a box-score comparison.
The Problem
A full postseason is about 85 games, and most players appear in four to six of them. That is a few hundred possessions against roughly 2,500 in a regular season. Run the ordinary RAPM fit on a sample that thin and the ridge penalty does exactly what it is supposed to do — it shrinks everyone toward zero and reports that nobody in the playoffs mattered.
The fix is to change what the model shrinks toward. Instead of pulling every player toward league-average, pull him toward his own regular-season rating. The question the fit answers then changes completely — from “what is this player worth, from scratch?” to “how did this player differ from his regular-season self?” That is a question a few hundred possessions can actually speak to.
The Method
Standard ridge penalises coefficients toward zero. Penalising toward a prior instead is a small change with a large consequence:
g is the playoff delta, and it is the output the page is actually about. It is reported separately from the level, because they answer different questions: the level is how good he was in those playoffs, the delta is how much that differed from the player he had been all year.
The Decisions
The prior is the three-year regular-season window ending that season. It covers the two prior seasons plus the one just finished, so nothing leaks in from the future.
The same season’s one-year rating is a tighter match to the player who actually walked into the playoffs, but it is noisier — and prior noise does not cancel. It lands directly in the delta and reads as “this player played differently in the postseason” when in fact the baseline was just badly measured. There is an ablation for exactly this: rerun with the one-year prior and compare the median absolute delta. If the deltas come out systematically larger, that is the noise showing, not players changing more.
The intercept and home-court terms start at zero and are fit free on playoff data. Playoff pace and playoff home-court advantage both differ from the regular season, and folding regular-season values into them would silently assert that they do not.
Rookies, and anyone below the regular-season possession cut, have no prior to shrink toward. They are recorded as such rather than being handed a zero: “no prior” and “measured at zero” are different claims and the output does not conflate them.
The Check
The obvious failure mode is that the playoff possessions add nothing at all and the deltas are pure shrinkage artefact. So the model is asked to prove otherwise, on held-out games, against two competitors:
| Predictor | What it is |
|---|---|
| Prior only | The regular-season rating alone, with the intercept and home court still fit free. Does the playoff data add anything at all? |
| Prior + delta | This model. |
| Plain ridge | Shrink to zero — what you would get with no prior. |
All three predict the same held-out stints, scored on possession-weighted RMSE. Where prior+delta fails to beat prior-only out of sample, the season’s deltas are flagged unreliable and that flag travels with the numbers into the table. Without it a reader cannot tell a measured +4 from a shrinkage artefact.
Worth being clear about what this means: the levels stay trustworthy in seasons where the deltas are not, because the level is mostly the prior. The check was built to be able to return the answer “playoff deltas are not measurable at this sample size” — and in some seasons it does.
Reading It
| Column | Meaning |
|---|---|
| Prior | His three-year regular-season RAPM going into the postseason. The bar he is being measured against. |
| Playoff | The rating fitted on that postseason’s possessions, anchored by the prior. |
| Δ RAPM | Playoff minus prior. Positive is a rise, negative is a fade. Split into offensive and defensive halves. |
| Poss | Playoff possessions behind the estimate. The single most important column for deciding how much to trust a row. |
| Box deltas | Playoff minus regular-season per-game points, assists, rebounds, true shooting and so on. Shown beside the model, and not inputs to it — they are there so you can see what the shift looked like in familiar units. |
All values are per 100 possessions, and a positive defensive number means points prevented.
Limitations
The sample is small and it stays small. Even with the prior doing the heavy lifting, a player with one first-round exit has a few hundred possessions behind his delta, and a deep run against three good defences is a genuinely different measurement from a five-game exit.
Opponent quality is not adjusted for beyond what the fit does naturally. RAPM controls for who was on the floor, so facing a great defence is partly handled — but a hard bracket and an easy one are not put on equal footing by any explicit term.
The prior is a three-year window, so a player who genuinely improved during the regular season walks in with a baseline that is partly two years stale. Some of what reads as a playoff rise is in-season improvement arriving late.
Strengths
It answers the question people actually ask — did he rise or fade — as a single measured quantity, rather than as a composite of hand-weighted box-score movements.
Level and delta are reported separately, so a good playoff run by a good player is not confused with a leap by an average one.
It says when it does not know. A season whose deltas fail the held-out check is labelled rather than published as if it were measured.