Playoff Rest

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, records series format. First rounds were best-of-five through 2001-02. Format adjusts the prior-round grind calculation but is not a separate model input.

logistic_grind_v2 superseded logistic_unreg_v1 on 2026-07-31, when entry_rest_diff (raw days of rest) was swapped for prior_grind_diff above.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.

A short prior series can indicate team strength as well as less workload. The first comparison keeps only teams that closed their own previous round early, then groups them by how long their opponent’s round lasted. This narrows the comparison but does not randomly assign opponents or eliminate differences in team quality.

16.7 percentage points separate the groups. Teams whose opponents went long were also slightly better by regular-season record, so that difference can confound the comparison.

Across second-round-or-later series with similar regular-season records, without holding the team’s own prior round fixed, the rates are 53.2% becomes 67.9% (62 series against 78), a gap of 14.7 percentage points.

For teams that went the distance themselves, the comparison reverses by 6.2 points. The association depends on both teams’ prior rounds.

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

What we cannot tell you: whether it is really fatigue. A team that needed seven games may also be weaker than its regular-season record suggests. These comparisons cannot separate that possibility from fatigue. The association does not fade game by game as a simple recovery explanation would predict.

“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.

In walk-forward evaluation over 30 held-out seasons (450 series, 1995-96 onward), the model produces lower probability-error scores than a constant historical home-court win rate.

Log loss and Brier score are both lower-is-better measures of whether a stated probability matches the outcome. Both penalise confident errors. The baseline always picks the home-court team and assigns its historical win rate as the probability.

The tests did not establish better winner accuracy. 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. The evaluation provides stronger support for improved probability estimates than for improved winner selection.

The dataset contains only a few hundred series. That limits precision, especially for round-level comparisons, and increases the risk of overfitting more complex models.

Accuracy differs by round in this sample. In Round 1, prior_grind_diff is always 0 because there is no prior round. The model still uses its other features, but its accuracy falls below the always-home-court rule. From Round 2 onward its observed accuracy is higher. This split alone does not isolate the grind term’s contribution.

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, a 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. Historical picks are walk-forward predictions: the target season is excluded from training.

A series’ hindsight figure comes from one model fitted across every covered season at once, including the one being predicted. This is an in-sample estimate and cannot establish predictive performance.

Hindsight exists for one reason: the model needs about ten seasons of prior history before its first walk-forward fit, so the earliest covered brackets have no such forecast. 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 and expected player availability.
• —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 changes are reflected only indirectly through the team's regular-season results.
• —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.