The MRI
2025–26 · Final

Model versus market

The model does not beat the market, and this page is not a card to bet. It is published because the comparison is interesting and because a system that only shows you its good weeks is not telling you anything.

The market is the better predictor

MRI 2.0Market
Mean margin error, 16,955 games (ESPN Bet) 9.40 pts8.86 pts
Mean margin error, 4,498 games (DraftKings) 9.52 pts8.81 pts
Picks the winner outright 71.6%73.0%

The two numbers correlate at 0.90, so most of the time they agree. Where they disagree it is usually the model that is wrong, and half a point of margin error is the whole reason there is nothing to bet here.

Against the closing line, by season

SeasonATS
202351.6%
202450.0%
202550.0%
202651.9%

Break-even at −110 is 52.38%. Four seasons, never close, no trend. On DraftKings over 2025–26 it is 48.3% across 4,498 games, which is worse still.

By size of disagreement

If the model knew something the market did not, the buckets where it disagrees most should be the ones that win. Against ESPN Bet's closing number:

GapBetsATS UnitsCLV
0-1 3569 51.3% -82.9
1-2 3309 50.1% -156.0
2-3 2755 50.1% -132.5
3-5 3779 49.9% -194.2
5-7 1993 51.4% -41.9
7-10 1109 51.6% -18.7
10-15 361 51.2% -8.6
15+ 79 68.4% +26.5

One bucket clears significance: gaps of 15 points or more went 54–25. There are 79 of those in 16,954 games, 42 of them in the first season tested, and reading them is the end of the story — they are November mismatches against opponents the model has barely seen. In one, the model favoured Mississippi Valley State by four where the market had Hawai'i by 25.5. It won the bet while being wrong by 22 points. The bucket is measuring the model's own failures, and it does not reappear on DraftKings.

The one faint signal

Betting the opening number rather than the close, mean closing line value is positive and grows with the size of the disagreement — the market tends to drift toward this model's side after the opener. That is a reason to keep measuring, not a reason to bet, and it is the same shape of hint football produced.

GapBetsATS UnitsCLV
0-1 884 51.4% -19.0 +0.08
1-2 846 49.9% -44.4 -0.02
2-3 763 48.2% -66.5 +0.11
3-5 1010 48.2% -88.3 +0.21
5-7 526 45.6% -74.6 +0.22
7-10 302 50.0% -15.1 +0.58
10-15 76 53.9% +2.5 +0.68
15+ 10 70.0% +3.7 +8.20

DraftKings openers, 2025–26.

Where they disagree now

No games in the next few days. This fills in during the season.