Updated Sep 7, 2026 · Week 1 board.
A panel of 64 models, not a single number
The system logs 1,653 individual picks across the 16 Week 1 games, out of 64 separately named models. A handful are flagship House NFL models. A couple of families start from simpler old-school ratings instead of this week's box scores. Roughly 50 more are research builds. Some run classic statistics. Some run newer machine-learning methods hunting patterns a person walks right past. One just blends other models' calls into a single number. Every pick is labeled a spread call or a moneyline call or a total call. Collapse that to one reading per model per market, keeping each model's latest, and this single game carries 63 distinct readings: 38 on the spread, 5 on the total, 20 on the moneyline.
One game, and the numbers do not agree with each other
Take those 38 spread readings, with San Francisco away and Los Angeles at home. The live market has Los Angeles favored by 3.5 points at all three books. Here's a sample of what the panel says about that.
| Model | Projected number | Side | Line | Odds | Book | Date |
|---|---|---|---|---|---|---|
| Our Main Model | 6.00 | home | 3.5 | -110 | bovada | 2026-08-21 |
| A Simple Reference Model | 1.06 | away | 3.5 | -105 | bovada | 2026-09-01 |
| A Preseason-Ratings Model | 4.58 | home | 3.5 | -113 | pinnacle | 2026-09-04 |
| A Model That Blends Several Others | 0.80 | away | 3.5 | +100 | pinnacle | 2026-09-05 |
| A Pattern-Matching Model | -2.06 | away | 3.5 | +100 | pinnacle | 2026-09-05 |
| A Curve-Fitting Model | -15.38 | away | 3.5 | -110 | betrivers | 2026-09-04 |
Thirty-five of the 38 spread models take San Francisco, the away side, at the market's 3.5. Three land on Los Angeles: Our Main Model, the preseason-ratings model, and the one that blends other models together. The projected numbers run from an outlier of -15.38, our curve-fitting model, a method famous for going haywire on small samples, up to 6.00 from Our Main Model. Averaging that scatter into a tidy headline throws away the real disagreement among 38 separate runs on one football game, on the same afternoon.
How a signal earns its way into a spread number
Every one of these models gets asked the same question in its own accent. Given what it's been fed, where should this line sit next to where the market has it? A box score tells you what happened in a game that already ended. Every model here gets judged one way, whatever method it runs on. Did it call the outcome before the outcome existed. That's how one game produces a 21.4-point spread in projections, from 6.00 down to -15.38, among models fed the same public information. Whatever a model watches is worth something only if it moved that model closer to the real final score across many games. That scoring hasn't happened yet for Week 1.
Where the published pick actually comes from
Our picks page lists one house pick per market here. On the spread, home side, line 3.5, projected number 6, edge 2.5 points (the line subtracted from our number), confidence 0.75 (the model's own self-rated odds), priced at -112. On the total, over, line 48.5, projected 50.29, edge 1.79 points, confidence 0.62. On the moneyline, home side, projected 0.66, edge 0.16, priced at -198. Those three projected numbers, 6 and 50.29 and 0.66, are Our Main Model's own numbers almost exactly. So the published pick traces back to one named model per market rather than a vote of the whole panel. On the spread it sides with 3 of the 38 models. The other 35 point the opposite way.
What has not been proven yet
Not one of the 64 models here, the one feeding the published pick included, has a graded result behind it. These are pre-kickoff numbers. What I can show you is which model produced which number, and how far apart the panel sits on a single game. Which model was right isn't something anybody can show you today.
The Receipts Drawer
The published pick is one model's call, not the panel's. On this spread that's a 3-model minority standing against a 35-model majority, with zero grades behind either group. What moves me off it is Our Main Model's call failing once Week 1 finals land, because a minority call that hits is worth far more to you than a majority that shrugs. So watch that one number, not the crowd around it.
Our published picks are at /picks/nfl, the market board they're priced against is at /odds, and the arithmetic for turning an edge into a bet size is at /desk/tools/ev. Once results start posting, you can check whether the published side beat the closing number at /desk/tools/clv. The wider view of every market and every model is at /desk.
