Skip to content

How accurate is the model?

We replayed the 2025-26 season day by day. Before every game, each player with 10+ games and 20+ minutes a night got a line — the half point over his season average so far — and the model's chance of going over it, using only games already played. Then the game graded it. A calibrated forecast lands on the diagonal: of everything it called 60%, about 60% went over.

Lines here are our own, not a sportsbook's; this measures the forecasts, not betting profit. Out of sample: tuned on 2023-24 and 2024-25, never on this season.

Lines graded
14,141
Points
Brier score, model
0.2429
coin flip 0.2500
Brier score, season hit rate
0.2515
model is 3.4% better
Model settings
half-life 5 · shrink 0.6
tuned by log loss

Forecast vs outcome

ModelSeason hit ratePerfectly calibrated
0%0%20%20%40%40%60%60%80%80%100%100%Forecast chance of the overHow often it went overModelSeason hit rate

Bins with fewer than 50 lines are left off. The hit rate's line stays flat: it barely tells overs from unders.

By bin

ForecastLinesModel saidWent overHit rate: said → over
0–10%0–––
10–20%0––18% 33% (3)
20–30%7628.0%13.2%26% 39% (77)
30–40%2,13636.8%35.2%37% 44% (1,264)
40–50%7,44245.2%44.5%45% 45% (7,601)
50–60%3,91353.7%52.3%53% 47% (4,881)
60–70%55063.3%63.6%62% 42% (306)
70–80%2471.7%62.5%71% 56% (9)
80–90%0–––
90–100%0–––

Do streaks go on?

The most striking active run per player (5+ games, at round marks near his average), graded the next game.

StreakStreaksWent onModel expectedHis season rate at the mark
Over PTS41661.1%60.3%53.2%
Under PTS44964.6%62.8%53.7%

Every stat, 2025-26

StatLinesBrier, modelBrier, hit rateBetter byLog loss
14,1410.24290.25153.4%0.6788
14,1410.24420.25022.4%0.6813
14,1410.24090.24732.6%0.6745
14,1410.21770.22232.1%0.6173
14,1410.24190.25234.1%0.6765