Rest advantage
The fatigue score: inputs, constants, historical comparisons, and limitations.
Keep in mind Historical rest groups also differ in home court and team strength. Their win rates do not isolate the effect of rest or predict an individual game.
On this page
How the score works
Each team receives a fatigue score based on its schedule before tip-off. A higher score represents more estimated fatigue. The difference between the two teams’ scores is the rest advantage, used to group games in the backtest.
baseLoad = recentWorkload + travel + roadSegment score = max(0, baseLoad × backToBack × altitude × density + freshness + overtime) restEdge = awayScore − homeScore (positive ⇒ the home side is fresher)
A difference under 0.5 is treated as neutral: 8,193 games, about one in six, that carry no rest claim in either direction.
Home court and rest: comparisons and baselines
Rest and home court are entangled, and the entanglement is structural. A visiting team has travelled by definition, so the fresher side is the home side in 27,404 of the 38,950 games with a measurable gap. Two of the eight terms, body clock and visiting altitude, are tied to playing away from home, and the schedule pushes the same way on its own: the visiting side is playing a second night in a row roughly twice as often as the home side, and nothing in that term knows which team is at home.
In about half the schedule the home team had travelled in too. In those games, the average rest edge falls by three quarters, while the home team still won 59.7% of those games, 60.0% of the ones where it had flown farther than its opponent, and 58.0% of the ones where it ended a road trip on a back-to-back. These groups all retain a substantial home win rate.
So a rest advantage cannot be read on its own. It has to be read against the venue it arrived with. Home teams win 59.9% of all 47,143 games and road teams 40.1%. The corresponding rested groups sit +1.3 percentage points above baseline at home and +2.3 on the road. These associations are small on both sides, and neither comes near the twenty points between the two baselines.
| RESTED TEAMWHERE IT PLAYED | GAMESCOUNT | IT WONWIN RATE | BASELINESAME SIDE, ALL GAMES | VS BASELINEPCT POINTS |
|---|---|---|---|---|
| Rested team at home · published | 27,404 | 61.2% | 59.9% | +1.3 |
| Rested team on the road · counted separately | 11,546 | 42.4% | 40.1% | +2.3 |
| No measurable gap · |RA| < 0.5 | 8,193 | — | — | — |
| Every completed game since 1985-86 | 47,143 | 59.9% HOME · 40.1% ROAD | ||
Model Results shows historical win rates, not the chance of winning a particular game. A rest advantage of 7+ means all games with a score gap of at least seven, not seven days of rest. Home teams are grouped by this minimum gap; road teams are pooled across all rest gaps. These rates use the current backtest; its group counts and threshold results are available in Model Results. The tables on this reference page retain their dated analysis samples.
The headline counts the rested home team. Rested visitors are reported separately: their pooled win rate is 42.4%, below 50% despite exceeding the road baseline. The headline therefore includes home-court advantage as well as rest.
| REST GAPAT LEAST | GAMESCOUNT | RESTED ROAD TEAM WONWIN RATE | VS 40.1% BASELINEPCT POINTS |
|---|---|---|---|
| any | 11,546 | 42.4% | +2.3 |
| ≥ 2 | 4,354 | 43.4% | +3.3 |
| ≥ 3 | 2,056 | 43.7% | +3.6 |
| ≥ 4 | 952 | 47% | +6.9 |
| ≥ 5 | 341 | 46% | +5.9 |
| ≥ 6 | 108 | 50% | +9.9 |
| ≥ 7 | 26 | 61.5% | +21.4 |
The last two rungs are 108 and 26 games in 41 seasons. Read those rates with caution because very large gaps have few examples.
The pooled ladder spans changes in home-court advantage. The era table compares rested visitors with the road baseline from the same period.
| ERASEASONS | GAMESCOUNT | RESTED ROAD TEAM WONWIN RATE | ROAD BASELINETHAT ERA | VS BASELINEPCT POINTS | BEST RUNGGAP · WIN RATE |
|---|---|---|---|---|---|
| All seasons · 41 | 11,546 | 42.4% | 40.1% | +2.3 | ≥ 4 · 47% |
| Last 10 seasons · 10 | 3,084 | 47.6% | 43.8% | +3.8 | ≥ 4 · 52.4% |
| Last 5 seasons · 5 | 1,638 | 49.3% | 44.7% | +4.6 | ≥ 3 · 52.9% |
The rested road team has climbed from 42.4% to 49.3%, which looks like rest coming to matter more. Most of it is not: the road baseline rose from 40.1% to 44.7% over the same span, because home-court advantage has weakened league-wide. What is left after subtracting that is the last column, and it has moved much less: +2.3 to +4.6 points.
In the last ten seasons, rested visitors with a gap of 4 or more won 52.4% across 231 games. That subgroup exceeds 50%, so the pooled road rate cannot support a claim that rested visitors always lose more often. This retrospective threshold comparison has not established a new prediction rule.
Folding home court into the score itself and letting the combined number pick was measured too: with a 3-point home bar it covers 96.5% of games at 59.7%, which is below simply picking the home team in every one of the 47,143. It also makes 1,386 road picks and loses 752 of them. This tested rule did not outperform the all-home baseline.
Full formula and constants
| Term | What it measures |
|---|---|
| Recent workload | Previous games, weighted by recency and final margin. |
| Travel | Distance between venues and estimated body-clock displacement. |
| Road segment | Consecutive road games beyond the initial allowance. |
| Back-to-back | A load multiplier based on time between tip-offs. |
| Altitude | Visiting altitude and next-night carryover. |
| Schedule density | Games across five windows compared with a normal pace. |
| Freshness | A discount for extended rest. |
| Overtime | Additional load from overtime in the previous game. |
Recent workload. Every game in the last 30-day window adds load that decays exponentially, so last night matters far more than last week. Each game’s cost is scaled down when it was a blowout, using final margin as a proxy for reduced workload. The score does not observe whether individual starters actually rested.
cost = 2.65 × e^(−0.52 × daysAgo) × blowoutFactor blowout = 1 − 0.25 × clamp((|margin| − 15) / 20, 0, 1)
Travel. Great-circle miles between consecutive venues over a 7-day window, log-scaled so additional miles add progressively less to the score. The itinerary assumes a trip home only when the next game is at home.
travel = 1.75 × ln(1 + miles / 1000)
Body clock. A charge for playing at least a 2-hour clock shift from home, resolved from each venue’s UTC offset. The model assigns a larger multiplier eastward and reduces the charge with nights in the new zone. These are retained assumptions; the Time Zones analysis did not validate their predictive value.
displacement = 0.88 × direction × max(0, 1 − nightsInZone / zonesCrossed) direction = 1.25 eastward, 0.85 westward
Back-to-back. Playing last night multiplies the load. The size of that multiplier depends on the real gap between tip-offs, because a 10:30pm game into a 7pm game is roughly 21 hours of recovery and the reverse ordering is 27.
b2b = clamp(1.38 + 0.02 × (24 − turnaroundHours), 1.3, 1.46)
Road segments, altitude, density, extended rest and overtime. Consecutive road games add 0.34 each after the first 2 are free. Visiting altitude (Denver, Utah, and Mexico City at 7,350 ft) multiplies by 1.29, and the following night at normal elevation by 1.06. Schedule density compares games played across five windows against a normal pace. Extended rest earns a discount that begins at 3 days and approaches −2.0. A prior game that went to overtime adds 0.5, or 1.0 for double overtime or more.
What changes when each term is removed
Measured 2026-10-04
The model always selects the home side when it identifies a rested-home game. Removing a term changes which games meet that threshold. Each term was neutralized in turn and the win rate recalculated against a baseline of 61.18% across 27,404 games.
Games selected by back-to-back and recent-workload terms but lost when each is removed have win rates above 63%. Removing travel drops 5,992 of the model’s calls, more than twice any other term. Those games won at 59.1%.
| TERM | GAMES ONLY IT FINDSCOUNT | THOSE GAMESWIN RATE | HEADLINE IF REMOVEDPCT POINTS |
|---|---|---|---|
| Travel | 5,992 | 59.1% | +0.34pp |
| Recent workload (decay) | 3,434 | 63.34% | −0.68pp |
| Back-to-back | 1,742 | 63.61% | −0.32pp |
| Road segment | 2,785 | 58.92% | +0.24pp |
| Altitude | 617 | 62.4% | −0.04pp |
| Schedule density | 709 | 60.37% | −0.07pp |
| Overtime | 154 | 64.29% | −0.08pp |
| Freshness | 148 | 60.14% | +0.01pp |
Removing travel or road segment raises the published rate while reducing coverage. The games they add win less often than the model’s overall 61.18% rate. Removing travel reduces net correct calls above a coin flip by 398 and drops 5,992 called games. Those are all dropped calls, including losses, not a count of winning predictions.
The whole model sits at 3,063 correct calls above a coin flip across 27,404 games. Freshness is the only term whose removal improves this net count, by ten calls. The ratified terms remain in place under the evaluation protocol in ADR 0006. Terms interact multiplicatively, so these figures do not sum to the total.
The counts in this section are stated against a coin flip rather than against the 59.9% home baseline used everywhere else on the site. That measures accuracy and coverage together, but it gives credit for home-court advantage too. It does not measure the model’s improvement over home court; the rested home group sits +1.3 points above that baseline, as shown on Model Results.
A separate out-of-sample fit found travel adds little independent information once the other schedule terms are known. The ablation table answers a different question: which games cross the fixed call threshold when travel is included. Changing coverage is not evidence that travel improves an independently fitted model.
Data sources and revisions
Schedules, scores and results come from the NBA’s own feeds. Overtime periods, tip-off times and neutral-site venues come from ESPN. The missing overtime input was restored on 2026-07-30; earlier model runs had treated it as zero throughout.
Arena coordinates are era-correct: Sonics games resolve to Seattle, not Oklahoma City, and the 2005-06 Hornets to their Katrina-season home. Distances are great-circle, not routed.
ESPN coverage begins around 2002, and its neutral-site flag only from 2013. Earlier seasons are scored by the same formula with those three inputs absent, which means a pre-2002 overtime count of zero denotes unknown, not “no overtime”. The trade was taken deliberately rather than restricting the model to a shorter span, and the cost is that a 1994 score and a 2024 score are not built from quite the same information.
Full limitations
The model reads schedules. It knows nothing about the teams playing.
- —No injuries, rotations or minutes played. A rested team missing two starters scores the same as a healthy one.
- —No team quality. The historical groups can differ in strength as well as rest, so their win rates do not isolate a causal rest effect.
- —No actual itineraries. Teams are assumed to fly venue to venue and only home when the next game is home. No public source records what they really did.
- —No load management. A star sitting a back-to-back is exactly the effect this model would want to capture, and it is invisible here.
- —Playoffs are excluded entirely. A fixed two-team series breaks the travel assumptions.
- —The 2019-20 Orlando bubble is excluded because it had one site, no travel, and no crowd. Pre-suspension games remain included. The condensed 72-game 2020-21 season is included, with greater density than the model's normal-pace anchors.
Evaluation protocol and model history
Every constant above was set by reasoning about the physical effect and reviewed before the backtest was run; none was fitted to maximise this win rate. Reusing evaluation results to choose constants can overstate performance. One constant has moved on measured evidence: the altitude multiplier was raised from 1.15 to 1.29 on 2026-08-02 to match altitude’s measured size against a back-to-back on final margin. That is a different target from the win rates on this page, and the change is recorded in ADR 0006. Using a different target does not make the shared historical games an independent test set.
In the overhaul recorded on 2026-07-30, nine fixes landed together and the published hit rates rose about a point. On games both the old and new model called, accuracy moved 0.15pp and the two picked the same team 98.8% of the time. The gain was almost entirely the new model declining 2,661 games the old one had called at below a coin flip. Most of the reported improvement came from selecting a different set of games.