Updated Sep 7, 2026 · Week 1 board.
Elo gives you a rating, not a reason
An Elo rating boils a team down to one strength number, turns two of those into a predicted margin, then updates after every result. What it never tells you is which yards or which plays moved it. Eight of the engines we point at Week 1 run on some version of that rating (Elo: one strength number per team). We pointed all eight at this game.
Some are tuned on preseason form, some on the rating alone, one pair Bayesian (it updates a little at a time) and one pair Linear (it updates by a fixed step). The table below carries the full breakdown. Same two teams every time. A different number every time.
A rating is an opinion with a number attached, and eight opinions came back.
Eight Elo-family rows, one Sunday-night game
New England at Seattle kicks off at 8:20 p.m. ET, Sep 9. Every Elo- and SHAP-named engine we carry got pointed at that one game.
The market on it is tight: Pinnacle, DraftKings and Bovada all have Seattle favored by 3.5 points with a 44.5-point total, checked between 4:52 p.m. and 6:07 p.m. ET, Sep 7. Against one settled number, the eight engines come back scattered.
| Model | Model number | Pick & line | Run on |
|---|---|---|---|
| House NFL Bridge - Preseason Prior (Elo) - Spread | 5.7004 | home 3.5, −108 | betmgm |
| Engine A: SHAP Explainable Elo Model (Linear) | −4.4547 | away 3.5, −110 | bovada |
| Engine B: SHAP Explainable Elo Model (Linear) | −4.4521 | away 3.5, −115 | draftkings |
| Engine A: Simple Elo Spread Model (Linear) | −3.7767 | away 3.5, −110 | bovada |
| Engine B: Simple Elo Spread Model (Linear) | −3.7801 | away 3.5, −115 | draftkings |
| Engine A: Simple Elo Spread Model (Bayesian) | −3.5849 | away 3.5, −110 | bovada |
| Engine B: Simple Elo Spread Model (Bayesian) | −3.5849 | away 3.5, −115 | draftkings |
| NFL Elo Power Ratings Spread (XGBoost) | 1.2344 | away 3.5, −115 | draftkings |
Read that number column carefully, because I nearly didn't. Nothing tells us these eight models share the same plus-or-minus convention, so comparing the raw numbers head to head is a guess dressed as a fact. What we can say is which side each model actually took. Seven of the eight are on the away side, New England, against the 3.5-point number.
The Preseason Prior (Elo) model stands alone on Seattle at that same number. That split is sourced. Seven one way, one the other, and that beats averaging a column whose sign nobody has checked.
What "SHAP" promises that is not in the data yet
SHAP stands for Shapley Additive exPlanations. It's a method for cutting a model's number into the pieces that built it. Done properly it reads like this: home turf plus 0.8, rest advantage plus 0.3, pass-rush mismatch minus 1.1, and those add up to the final number. Two models in the table above are literally named SHAP Explainable Elo Model.
Neither one shows me a piece-by-piece breakdown. Each is a single number, a side, a line and a date it ran, the exact shape of every other model up there. So the name is promising an explanation that isn't in front of us this week. If that breakdown lands later, I'll say so right here.
Where the proof still has to start
There's no track record underneath any of this. Across every house pick we've graded site-wide, the count sits in two buckets: 6,119 we can't grade at all, and 1 marked a loss. Not one win yet. That tally covers the whole site rather than these eight models, so no separate accuracy number exists for any of them, Elo or SHAP-named or otherwise.
What I can stand behind this week is the number each model produced, the side and line it took, and the plain fact that the SHAP-named ones don't show the breakdown their name advertises.
The Receipts Drawer
Eight models, one rating idea, eight different numbers, and two of them landing on the identical figure: both Bayesian Simple Elo Spread engines print −3.5849 on this game, independently. Agreement that exact is rare enough here that I stopped and checked it twice. Everything else spreading out the way it does is a reason to shop this game before you bet it, not a reason to skip it. The SHAP label is a promise the numbers don't keep yet, so don't pay extra confidence for a model just because its name says it explains itself.
Go read the receipts yourself. Every published Week 1 pick sits at /picks/nfl. The live market is at /odds. The player side of this same game runs off the Week 1 projection board, which gives Jaxon Smith-Njigba 18.92 points in PPR scoring, third among receivers, and Drake Maye 16.26, thirteenth among quarterbacks: /fantasy/nfl/rankings.
How our spread models get built is at /blog/nfl-betting-model-explained, and why a SHAP-named model still does not show its work is at /blog/glass-box-vs-black-box-betting-models. Watch for a rerun that finally attaches a real breakdown to those two. Until it does, treat every model in that table the same way: a number, a side, and nothing else.
