Betting

That is the expected result and worth stating first. A closing spread is the sharpest number in sports, and a rating built from final scores is not going to out-argue thousands of people with money at stake. Anyone whose model beats closing lines by a point and a half is either wrong about their backtest or should not be publishing it.

Against the number the market opens at

Openers are a different question, because they are posted before the market has digested anything. Here the picture is better but not conclusive: 52.7% over 1,975 bets (+12 units) across 2023–2025, the only seasons for which opening lines exist.

That carries p = 0.4111. It is not significant, and the best individual bucket does not survive correction for the number of buckets examined. Treat it as a hint.

The one clean signal

Closing line value — whether the market moves toward your side after you bet — is the honest early indicator, because it converges in hundreds of bets where win-loss records need thousands. It rises steadily with the size of the disagreement:

A dose-response pattern across ordered bins is much harder to produce by chance than one good bucket, and it is positive in all three seasons. It says the market tends to drift toward this model's side after the opener — which is what genuine information looks like, and is the reason to keep measuring rather than to start betting.

How much more would it take to know

To confirm a true 53.5% edge at conventional power would take roughly 12,000 bets — about seventeen seasons of betting every game. Even a 56% edge would need close to two full seasons. The honest conclusion is that this question cannot be settled from history. It can only be settled forward, which is what the card below is for.

The record so far

Two records, and only one of them is a track record.

The season, reconstructed

For every game already played, what the model would have said using only the ratings from before that week. Nothing from the week it is predicting leaks in, but it was computed afterwards, and a record computed afterwards is a backtest however carefully it is done.

That is 1.3 points worse than the market (±0.4). Over the seventeen archived seasons the model averaged 13.0 points and picked 73.8% winners; a September sample has far less to go on, because the ratings lean on last year until this year's results replace them. A size of 1.00 means the lines are the right size; below it, they are too extreme — a favourite laying more points than it wins by.

Where the model disagreed with the opening number by 3 points or more, the board's own rule: 139–148–7 against the number (48.4%; break-even is 52.4%), -21.6 units, mean closing line value +0.7. That is 294 bets, far too few to say whether the disagreements mean anything.

The forward log

Started 2026-09-19. Each flagged game is written down before kickoff — model line, the market's number at that moment, the side — and never edited afterwards. Games are flagged on the gap between the model and the opening line, but a pick is graded against the number the market was showing when it was written down, at −110, because that is the price a bettor could actually have had. Every line is the home team's expected margin, and Edge is the model's minus the logged one. In 146 of the 289 picks the line had already moved by the time the pick was logged; the edge shown is against the line as it then stood. 48 of the picks were logged before the preseason prior began to use roster talent and returning production, so they were made by the earlier model; they stay as written.

124–134–6 on 264 graded of 289 logged (-21.3 units, mean closing line value +0.2).

Showing 40 of 289.

This week

Tracked, not recommended. These are the largest disagreements with the opening number. They are published so the record accumulates in public, including the losing weeks.