AVRCFB Model

The Model

Every published prediction, scored against the result.

Season
2026
Graded
4
Straight up
75%
Straight up
75.0%
3/4 winners called
Margin MAE
17.45
RMSE 18.4 · 4 games incl. FCS
Bias
-5.65
mean signed error
Correlation
-0.194
predicted vs actual r
Brier score
0.215
log loss 0.655
80% coverage
50%
target 80%

How close it gets

4 graded games · 2026
Every graded game

Predicted vs actual margin

-42-42-21-2100+21+21+42+42Wk 0 UNC @ TCU: model +19.2, actual -5Wk 0 SJSU @ USC: model +38.0, actual +16Wk 0 NCSU @ UVA: model +11.7, actual +26Wk 0 JXST @ NDSU: model +16.7, actual +26PEARSON r-0.194ACTUAL MARGINPREDICTED MARGIN
Error distribution

Residuals

21-24.5 to -17.5: 2 games3.5 to 10.5: 1 games10.5 to 17.5: 1 games-21-70+7GAMESNormal μ -5.7 · σ 20.3ACTUAL − PREDICTED
Season shape

By week

By predicted spread

Blowouts vs coin flips

|spread|straight-up %margin MAE
7-14 (1)
100%
14.3
14-21 (2)
50%
16.8
21+ (1)
100%
22.0

Does it know what it knows?

Brier 0.215
Win probability

Calibration

0%0%25%25%50%50%75%75%100%100%Predicted 79%, won 100.0% of 1 gamesPredicted 89%, won 50.0% of 2 gamesPredicted 100%, won 100.0% of 1 gamesOBSERVED WIN RATEPREDICTED WIN PROBABILITY
Prediction intervals

Interval coverage

80% interval50.0%vs 80%-30.0pp
90% interval100.0%vs 90%+10.0pp
95% interval100.0%vs 95%+5.0pp

The receipts

7 games

Worst misses

actual − model
Wk 0
UNC15@TCU10
model +19.2 · actual -5
-24
miss
Wk 0
SJSU26@USC42
model +38.0 · actual +16
-22
miss
Wk 0
NCSU8@UVA34
model +11.7 · actual +26
+14
miss
Wk 0
JXST7@NDSU33
model +16.7 · actual +26
+9
miss

Closest calls it got right

pregame favorite
Wk 0
NCSU8@UVA34
79%
fav
Wk 0
JXST7@NDSU33
88%
fav
Wk 0
SJSU26@USC42
100%
fav

Known limitations

Written down here, because a scorecard that only lists what went well is advertising.

  1. 01

    Intervals under-covered: 50% against an 80% target. The served σ is too tight.

  2. 02

    Projected totals are weak (MAE 12.3). The margin is what this model does well.