Game totals: why we publish a range and not a pick
For every game on the board we publish a range of plausible final scores with the market’s posted total marked on the same axis. We do not publish an over/under lean. That is a deliberate choice made after two rebuilds failed, and this page explains what the range claims, what it refuses to claim, and how you can check whether it is honest.
What the band claims
The claim is coverage, and it is falsifiable. When we draw an 80% band, roughly four finals in five should land inside it over a season. Held out of training, our bands covered 50.1 / 81.1 / 89.9 percent of NFL games against a nominal 50 / 80 / 90, and 50.5 / 81.1 / 91.4 percent of FBS-vs-FBS college games. CFB runs slightly wide at the 90% level, which is a small miscalibration we would rather name than round off.
Why there is no over/under call attached to it
Because we measured what our number adds to the market’s and the answer was: almost nothing. Blending our total into the posted line, with the blend weight chosen out of sample, optimises at a weight of 0.05 to 0.10 and improves the score by 0.0001. The market is also sharper than us outright. Both of those are expected — game totals are roughly half as predictable as margins for everyone, and the market explains only about 9% of the variance in totals — and neither is a reason to withhold a calibrated band.
What they are a reason to withhold is the gap between our middle and the posted line. A difference that carries no information is not an edge, and printing it invites you to read one. So we never show it, in any units, under any label.
The failure this replaced
Earlier versions of this site did publish a totals lean, and it was worse than wrong — it was hollow. Our model’s total moved much less than the market’s did, so whichever side of the line it landed on was mostly a mechanical consequence of where the line was set, not an opinion about the two teams. A crude rule that never looked at the model at all — take the under when the posted total is above the league average, the over when it is below — reproduced 78% of our published NFL sides. On one live board it matched all sixteen.
Two independent rebuilds, one a ridge regression on opponent-adjusted efficiency and one a drive-level Monte Carlo, were built without knowledge of the earlier work. Both landed in the same place and both failed the same pre-registered check. That agreement is why we stopped treating it as a bug to fix.
The probability we could show you and do not
A calibrated distribution can produce the probability that a game finishes over the posted line. It would be a true number. We still do not publish it, and this page is the only place we will even discuss it, because a probability attached to a priced proposition is a pick — it names a side, and the fact that it wears a percentage sign rather than the word “over” does not change what a reader does with it.
This is enforced structurally rather than by restraint. The data behind every totals surface ships a fixed set of quantiles rather than the parameters of a distribution, so nothing downstream — including anyone reading our public data directly — can compute that probability from what we publish.
How to check us
Coverage is tracked on /calibration. That row is currently marked as documented rather than live: the figures above come from held-out historical games, and we do not consider a coverage claim earned until the production board has graded at least 250 real games in each sport against the bands it actually published, frozen before kickoff. Until then the honest label is “measured in backtest”, and that is the label it carries.