Playoff predictions

One row per playoff series. The model outputs the probability that the home-court team wins the series — not the higher seed, and not a named favourite. Home court is the reference because it is unambiguous in every era and needs no bracket lookup, which seeding does.

P(home-court team wins the series)   →   ≥ 0.5 predicts them, otherwise the opponent

This is not a playoff version of the fatigue model, and it is worth being blunt about that because the site used to imply it was. It shares a philosophy with the rest model — that rest is a measurable input nobody prices — but it shares no code, no constants and no feature definition. Playoff games are deliberately excluded from the fatigue model itself: a fixed two-team series breaks its travel assumptions, since the opponent never changes and the itinerary is known in advance.

Weights are standardized logistic coefficients, so they are comparable to each other directly. All four are positive: every dimension of home-court advantage pushes the probability the same way, which is why the model almost always picks the home-court team and why its picks are hard to distinguish from that rule.

A fifth column, is_best_of_7, is stored on each series — first rounds were best-of-five through 2001-02, and a shorter series is more random — but it is not fed to the model. This page previously listed it as an input; it is not one.

logistic_grind_v2 superseded logistic_unreg_v1 on 2026-07-31, when entry_rest_diff (raw days of rest) was swapped for prior_grind_diff above. The v1 prediction rows are retained rather than overwritten, so older predictions stay auditable.

The Playoff Rest page shows teams winning far more often when their opponent came out of a long previous round. Fair objection: you earn a short series by being good, so maybe the fresh team just wins because it was better all along. The reason that does not cover it is that how long your opponent’s last series went is not up to you — it was decided by two other teams. So hold your own last round fixed at a quick close, and let only their side vary.

16.5 points, from something you did not control. But read the last column honestly: the teams whose opponents went long were also slightly better on record, so part of that gap is quality rather than exhaustion.

So widen back out to every second-round-or-later series — no longer holding your own last round fixed — and keep only the evenly-matched ones, where neither side has a real record advantage to hide behind: 53.2% becomes 67.9% (62 series against 78), a gap of 14.7 points. It barely shrinks.

And running it the other way — when you are the one who went the distance — moves it 6.2 points the wrong way, which is the signature of a differential rather than of long series being bad in the absolute.

The same thing counted a second way, by the layoff into Game 1 rather than by the previous round’s length — rounds 2+:

What we cannot tell you: whether it is really fatigue. A team that needed seven games to get past someone has also just shown it is worse than its record said — and this data cannot separate “worn down” from “not as good as we thought.” Game-by-game the edge does not fade the way tiredness should, which cuts against the fatigue reading. The effect is solid; the reason for it is arguable, and we would rather say so.

“Closed it early” means a team won its previous round within one game of a sweep; “went the distance” means it needed the last game or the one before it. Grind is counted as games beyond a sweep rather than as raw games played because 136 of 320 first rounds in this record were best-of-five, where five games means a team went the full distance rather than closing early.

The honest result splits in two, and only one half is good. Measured over 30 seasons predicted in advance (450 series, 1995-96 onward), the model produces materially better-calibrated probabilities than the base rate — it knows the difference between a lopsided matchup and a near coin flip.

Log loss and Brier score are both lower-is-better measures of whether a stated probability is honest: a model that says 90% and is right nine times in ten scores well, and one that says 90% and is right six times in ten does not. The base rate is the simplest possible competitor — always predict the home-court team, at the historical rate they win.

On accuracy that competitor is just as good. Across the same seasons the model beat it, tied it, and lost to it 11/13/6 times, and the confidence interval around the model’s accuracy contains the base rate outright. So the correct reading of this page is: use the probability, ignore the pick.

The dataset is small by the standards of any modelling problem: a few hundred series across the covered seasons, against roughly forty-six thousand regular-season games. That is the central constraint here, and it is why the model stays deliberately simple rather than reaching for something expressive enough to overfit — and why no amount of further work turns this into a strong classifier.

The pooled accuracy row above hides a split, and pooling is what produced the earlier “no real edge” reading. In Round 1, prior_grind_diff is 0 for every series by construction — there is no prior round to have been ground down by — so the model knows nothing the always-home-court rule does not, and loses to it. From the second round on there is a grind to read, and it wins there.

210 series is not many, so one pooled number is not proof. Season by season, from the second round on, the model beat the always-home-court rule in 11 seasons, tied it in 16, and lost to it in 3 — the paired, same-brackets-same-seasons comparison this claim actually rests on.

A series’ pick comes from a model trained only on seasons that had already finished when that series was played. That is a real forecast, and it is the only figure treated as evidence anywhere on the site.

A series’ hindsight figure comes from one model fitted across every covered season at once, including the one being predicted. It already knew the answer, so it flatters itself and is not evidence of anything.

Hindsight exists for one reason: the model needs about ten seasons of prior history before its first honest fit, so the earliest covered brackets have no forecast at all. For those seasons the hindsight figure is the only number that exists, and the page labels it as such. For every later season the product page shows the forecast beside the hindsight figure, each labelled, so the two are never mistaken for each other.

• —Injuries, which decide playoff series more often than any feature in this model.
• —Matchup and style. A team built to beat one opponent and not another is invisible to win percentage and seeding.
• —In-series adjustments. Coaches change rotations and schemes between games; the model predicts once, before game one.
• —Roster change between the regular season and the playoffs — a deadline acquisition counts only through whatever win percentage it produced.
• —Seeds are derived from win-percentage rank rather than read from an official bracket, so they can disagree with the published seeding in tiebreak situations.
• —A probability near 0.5 is the model saying it does not know. It is not a lean worth acting on.