Model vs Market

What happened when we disagreed

Every disagreement this season, graded against the spread, win or lose - the whole record rather than a selection from it, updated every Sunday whether it flatters us or not. A game is only graded on the rating we held before it kicked off, so the first few weeks of a season never appear: there is no honest way to say what we would have predicted before we had seen anything. That makes the count small in October and worth something by December. The three-season backtest below is the better evidence until then.

What this page is, and is not

It is not a tip sheet. We tested that before building it. Across three seasons and 799 games, betting every three-point disagreement with the closing line won 53.2% against a 52.4% break-even - which looks like an edge and is not: the odds of a run that good by chance are about one in three. Totals did better at 57% on a six-point gap, but eight thresholds were tried and one landing there is what luck looks like. Anyone selling you a confidence percentage built on numbers like those is selling you the percentage.

What the model demonstrably does have is information the opening line lacks. When we disagree with the opener, the market moves toward our number 64% of the time, and 70% when the disagreement is three points or more - both at odds of less than one in a thousand against chance. That is the usual evidence that a model knows something real, and it also explains the paragraph above: the information is genuine, and by kickoff the market has already priced it in. One limit worth stating - an opener posted after the previous weekend has seen the same games our ratings have, so part of that agreement is the two of us reading the same news.

Where the model is weakest, stated plainly. It knows what has happened on the field this season and nothing else - no recruiting, no roster talent, no preseason expectation, no injuries, no coaching. The market knows all of it. So in October you will see disagreements of twenty points and more, and they are mostly the model being young rather than the model being right: a team that is unbeaten against a soft schedule rates highly here long after the market has worked out what it is. Those gaps narrow as the season fills in. A wide disagreement in week six is a question worth asking; the same gap in week twelve is worth more.

So: the disagreement is shown, every one of them is graded, and the record stands in public. Our win percentage is the model's own number from the rating gap, not a market price with the margin stripped out, and it is the same arithmetic for every game whether we like the answer or not. Lines are DraftKings where we have them and Bovada otherwise, as held when this page was built.