Model evaluation

Performance

One record: every graded game since 2022-23, one forecast each, from a model that had not seen the game. Most of these forecasts were produced after the fact by the walk-forward evaluation; the rest were saved by the daily pipeline before tip-off.

How both are produced: Methodology.

The record in brief

Oct 18 – Nov 8

Measure	Result	What it means
Games evaluated	5,390	Regular season and playoff games from 2022-23 to 2026-27, each predicted by a model trained only on earlier seasons. 2026-27 is in progress.
Winners picked	65.1%	The team with the higher projected score won this often.
Home team won	55.6%	The baseline: picking the home team every time would have been right this often.
Margin error	11.0 pts	Average distance between the forecast margin and the final margin.
Total points error	15.0 pts	Average distance between forecast and actual combined score.
Team score error	9.4 pts	Average miss on a single team's score. Half of team scores land within 8 points.
Brier score	0.217	Average squared error of the home win probability. Lower is better; always saying 50% scores 0.250.

This season so far

2026-27, in progress

Share of games so far in which the forecast winner won, against the share won by the home team. Lines start once there are 10 games; early points move a lot.

98–48

146 graded games · 67.1% winners

Average margin miss: 11.7 pts
Brier score: 0.208
Home team won: 51.4%

Last 10 games

Nov 7, ORL at GSW: correct
Nov 7, HOU at PHX: missed
Nov 7, SAS at POR: correct
Nov 8, PHI at CHI: correct
Nov 8, LAL at DET: missed
Nov 8, UTA at MIN: correct
Nov 8, BKN at DAL: correct
Nov 8, BOS at MIL: correct
Nov 8, NOP at LAC: correct
Nov 8, MIA at SAC: correct

Winner calls

Every graded game, sorted by which side the forecast picked and which side won. The diagonal is the hits.

	Home won	Road won
Picked home	2,25142%	1,13521%
Picked road	74514%	1,25923%

The forecast picked the home team in 63% of games and was right 66% of the time when it did; road picks were right 63% of the time.

Last 30 graded picks

Oldest to newest: 26 right, 4 wrong. Each square opens the game.

Accuracy by season

Share of games where the forecast winner won.

● Regular season ■ Playoffs (about 84 games a season) Home team won (the baseline of always picking the home team).

Data table

Season	Regular season	Games	Playoffs	Games	Home team won
2022-23	62%	1230	61.9%	84	58.1%
2023-24	64.6%	1230	59.8%	82	54.6%
2024-25	67.3%	1225	57.1%	84	54.6%
2025-26	67.3%	1225	65.5%	84	55.5%
2026-27	67.1%	146	–	0	51.4%

Season	Games	Winners	Margin err.	Brier
2022-23	1,314	62.0%	10.3	0.231
Playoffs	84	61.9%	11.4	0.243
2023-24	1,312	64.3%	11.3	0.223
Playoffs	82	59.8%	12.8	0.234
2024-25	1,309	66.7%	11.0	0.211
Playoffs	84	57.1%	12.5	0.233
2025-26	1,309	67.2%	11.4	0.206
Playoffs	84	65.5%	12.9	0.227
2026-27 (in progress)	146	67.1%	11.7	0.208
All	5,390	65.1%	11.0	0.217

Accuracy rose as the training history grew: the 2022-23 model had one earlier season to learn from, the 2025-26 model had four. Playoff samples are small (about 84 games a season), so their year-to-year swings are mostly noise. The 2026-27 row covers only the games played so far; it is a small sample.

Every game, season by season

Orange: the forecast winner won. Grey: missed. Playoffs after the gap.

2022-23 815 / 1,314 · 62.0%

2023-24 844 / 1,312 · 64.3%

2024-25 873 / 1,309 · 66.7%

2025-26 880 / 1,309 · 67.2%

2026-27 98 / 146 · 67.1%

Hover or tap a dot to see that game.

Beyond the overall record

Team Predictability

Which teams' games this model forecasts well and which it misses: average margin miss with intervals, bias, variability, expected wins and Brier score for all 30 teams, by season.

Surprise Games

The results furthest from the forecast, ranked two ways: winners the forecast gave the lowest chance (this week: MIN at 13%), and the biggest misses on the final margin.

What moves the forecasts

How explanations are checked

By input group

Largest single inputs

Average size of each input's push on the projected margin, in points, across the 14 forecasts whose model could be verified (mostly production model saved 2026-11-07 and walk-forward 2026-27). Group values add up the inputs in the group, so they can exceed the largest single input.

Are the win probabilities honest?

Games grouped by the forecast's confidence in its pick.

Dot size reflects the number of games in each confidence band. Points above the dashed line mean the favorite won more often than forecast.

Calibration table

Confidence band	Games	Average forecast	Picks that won
50–55%	948	52.6%	53.4%
55–60%	1,108	57.2%	58.2%
60–65%	1,072	62.3%	61.9%
65–70%	838	67.4%	66.7%
70–75%	572	72.3%	74.1%
75–80%	466	77.4%	81.8%
80–85%	270	82.3%	84.4%
85%+	116	87.7%	88.8%

Reading the chart

If a 70% forecast is honest, teams given 70% should win about 70% of the time. Each band's points sit close to the diagonal, with the favorite winning slightly more often than forecast in the 70–80% bands.

Calibration note. Most of these probabilities come from the walk-forward evaluation's own calibration step. Upcoming forecasts use a probability curve fitted earlier on a much smaller set of games, so this chart describes the method, not a guarantee about any single percentage.

Predicted against actual margins

Average home margin, grouped by the forecast margin.

● Average predicted margin ○ Average actual margin.

Data table

Predicted range	Games	Avg predicted	Avg actual
Road by 10+	205	-12.1	-12
Road by 6–10	512	-7.3	-7.1
Road by 3–6	636	-4	-3.8
Road by 0–3	651	-1.4	-0.6
Home by 0–3	1,076	1.9	1.8
Home by 3–6	891	4.9	4.1
Home by 6–10	766	8.3	8
Home by 10+	653	13.6	13.2

What it shows

On average, games land close to where the forecast put them: when the model favors a team by a given amount, that team's average margin is in the same range. Individual games scatter widely around those averages.

Predicted margins have a standard deviation of 6.8 points against 15.5 for actual margins. A forecast is an expected outcome, so it is always narrower than real results, but it does mean a 20-point win is almost never forecast as one.

How big are the misses?

Distribution of errors across all graded games. Positive values mean the home side (margin) or the game (total) finished higher than forecast.

Margin error

Data table

Range (pts)	Count	Share
-40 to -36	30	0.6%
-36 to -32	37	0.7%
-32 to -28	65	1.2%
-28 to -24	107	2.0%
-24 to -20	146	2.7%
-20 to -16	255	4.7%
-16 to -12	337	6.3%
-12 to -8	438	8.1%
-8 to -4	577	10.7%
-4 to 0	608	11.3%
0 to 4	682	12.7%
4 to 8	569	10.6%
8 to 12	505	9.4%
12 to 16	365	6.8%
16 to 20	246	4.6%
20 to 24	176	3.3%
24 to 28	103	1.9%
28 to 32	63	1.2%
32 to 36	28	0.5%
36 to 40	53	1.0%

Values beyond the outer bins are counted in the first or last bin. Total: 5,390.

Total points error

Data table

Range (pts)	Count	Share
-50 to -45	26	0.5%
-45 to -40	37	0.7%
-40 to -35	69	1.3%
-35 to -30	116	2.2%
-30 to -25	163	3.0%
-25 to -20	275	5.1%
-20 to -15	369	6.8%
-15 to -10	455	8.4%
-10 to -5	498	9.2%
-5 to 0	592	11.0%
0 to 5	552	10.2%
5 to 10	534	9.9%
10 to 15	474	8.8%
15 to 20	364	6.8%
20 to 25	294	5.5%
25 to 30	212	3.9%
30 to 35	139	2.6%
35 to 40	89	1.7%
40 to 45	60	1.1%
45 to 50	72	1.3%

Values beyond the outer bins are counted in the first or last bin. Total: 5,390.

Team score error

Data table

Range (pts)	Count	Share
0 to 2	1,082	10.0%
2 to 4	1,440	13.4%
4 to 6	1,360	12.6%
6 to 8	1,232	11.4%
8 to 10	1,129	10.5%
10 to 12	961	8.9%
12 to 14	825	7.7%
14 to 16	723	6.7%
16 to 18	558	5.2%
18 to 20	432	4.0%
20 to 22	341	3.2%
22 to 24	219	2.0%
24 to 26	143	1.3%
26 to 28	112	1.0%
28 to 30	223	2.1%

Values beyond the outer bins are counted in the first or last bin. Total: 10,780.

36% of team scores land within 5 points of the forecast and 62% within 10.

Performance by context

All graded games split three ways. Every row shows its own sample size.

By predicted margin

Group	Games	Winners	Margin err.	Brier
Favored by 0–3	1,954	54.9%	11.1	0.247
Favored by 3–6	1,470	63.0%	11.2	0.234
Favored by 6–10	1,167	72.2%	10.9	0.201
Favored by 10+	799	83.7%	10.8	0.137

By which side was favored

Group	Games	Winners	Margin err.	Brier
Home favorite	3,386	66.5%	11.0	0.212
Road favorite	2,004	62.8%	11.1	0.226

Regular season and playoffs

Group	Games	Winners	Margin err.	Brier
Regular Season	5,056	65.4%	10.9	0.216
Playoffs	334	61.1%	12.4	0.234

Close forecasts are close to coin flips by design; accuracy rises with the size of the predicted margin. Road favorites are picked correctly less often than home favorites.

External benchmark: market lines

The model does not use betting lines. Where a betting line was stored for a game, it is compared here as a reference point. Lines are available for only 41 games (Apr 22 – Nov 8), far too few to draw conclusions.

Average error, 41 games	Model	Market line
Margin	10.7	10.4
Total points	13.8	12.0
Same favorite as the line	37 of 41