The market tells you what the line is. elo shap tells you why — and the dispersion number (how much books disagree) tells you whether to trust it.
When 71 books post the same number, the market has spoken with one voice. When they don't, the dispersion tells you the market is uncertain — and SHAP on an Elo model will show you exactly which inputs are fighting each other.
What is elo shap?
elo shap applies Shapley values (fair credit for each factor) to an Elo rating. It breaks a single rating number into the exact contribution of each input — opponent strength, rest, injuries, travel — so you can see why the model thinks a team is favored. A raw Elo number is a summary. SHAP unpacks it. You get a receipt. Opponent quality adds points. Rest adds points. QB injury subtracts points. Travel subtracts points. They add up to the rating. No black box.
What does the market consensus prove?
Our odds feed captures spreads across dozens of books. When every book posts the same number, the market has spoken with one voice. When they don't, the dispersion tells you the market is uncertain.
Look at the NFL slate from August 21, 2026. The Packers @ Broncos shows a home spread of -7.5 to -6 across 71 books with only 1.5 points of dispersion . That tight agreement means the market sees one clear story — Denver at home, Green Bay on the road, no QB drama. An elo shap breakdown for that game would concentrate most of the rating gap on home-field and roster quality, with tiny contributions from rest or weather.
Contrast Jets @ Steelers: home spread -3.5 to +2.5 across 71 books, 6 points of dispersion . A 6-point range means books genuinely disagree. SHAP on that game would show the rating drivers splitting — maybe rest advantage pulls one way, QB uncertainty pulls another, travel pulls a third. The model cannot decide because the inputs conflict. That dispersion is your signal to either pass or size down.
The same pattern holds across the board. Chiefs @ Buccaneers at -5.5 with 0 dispersion across 71 books — consensus. Commanders @ Lions at -4.5 to -1.5 with 3 points of dispersion — uncertainty. Saints @ Rams at +1.5 to +3 with 1.5 dispersion — mild disagreement. Every dispersion number is a proxy for how cleanly SHAP can attribute the Elo gap.
How do you read a SHAP waterfall for a game?
A waterfall (step-by-step breakdown) shows how the rating gap got built. Imagine the Elo model outputs a 1400 rating for Team A and 1520 for Team B. The 120-point gap implies roughly a 7-point spread. The waterfall shows the major drivers: base league average, opponent strength, rest advantage, QB status, home field, travel, recent form, injuries. The sum equals the gap. Now you know: if the QB clears, the gap shrinks by the QB factor. If the travel factor is overstated, that's another lever. You have actionable levers, not just a number.
Where does the market validate the breakdown?
The dispersion data is the ground truth. Games with 0 dispersion across 71 books (Bills @ Browns at -3 , Falcons @ Colts at -3.5 , Ravens @ Vikings at -3.5 to -3 with only 0.5 dispersion ) are the ones where SHAP produces clean, concentrated attributions. One or two factors dominate. The model agrees with the market.
Games with higher dispersion (Panthers @ Jaguars at -2.5 to +1.5, 4 points ; Giants @ Dolphins at +2.5 to +3.5, 1 point ; Bears @ Bengals at +1.5 to +2.5, 1 point ) produce SHAP waterfalls where credit is scattered. No single factor explains the rating. That is not a model failure — it is the model correctly reflecting a messy reality.
College football: same method, different consensus depth
CFB games in our data show two tiers of coverage. Some games reach 71 books: North Carolina @ TCU at -7.5 and San Jose State @ USC at -38.5 both show 71 books with zero dispersion. Other snapshots of the same games show 38 books: North Carolina @ TCU at -8.5 and San Jose State @ USC at -38 . The line moved a half-point as more books reported. SHAP still works — it tells you what the Elo model sees — but the dispersion proxy is weaker when fewer books report. For bettors, this means: trust NFL elo shap attributions more than CFB. In CFB, use SHAP to generate hypotheses, then verify against injury reports and practice news. In NFL, the market has usually priced the same information, so SHAP attributions align with line movement. You can track that alignment on the NFL picks page where model CLV (closing line value) is logged against the close.
How do you build your own elo shap pipeline?
You don't need a research lab. Pick features — opponent-adjusted efficiency, rest, travel, QB status, weather. Train an Elo-style rating. Run SHAP on every game in the backtest. Check whether games where SHAP concentrated on one factor produced better CLV than games where SHAP scattered. Our CLV variance study shows that single-driver games (QB injury, extreme rest gap) hit CLV targets more often than multi-driver games. The market dispersion data confirms it: low-dispersion games are the ones where the model and the market tell the same story. You can also learn the Elo math in our NFL betting model guide and the spread mechanics in how NFL spreads work.
What would change our mind?
If a game shows 0 book dispersion across 71 books but SHAP splits the rating across five equal factors, the model is hallucinating structure the market doesn't see. Conversely, if a game shows 6-point dispersion but SHAP concentrates 90% on one factor, the model is overconfident. Either mismatch means the feature set or the Elo formulation needs revision. Watch the Jets @ Steelers dispersion (currently 6) — if it collapses to 1 by kickoff and the line moves to -3.5, SHAP's QB-injury attribution was right. If it stays at 6, the market genuinely doesn't know, and SHAP should reflect that uncertainty.
Sources
- line-05831e33d31b0aacde409a089a422e6b — Braves @ Brewers home spread -1.5 across 9 books
- line-7bd4a1cf68fa48420e0ea7b5af4d4cc0 — Jets @ Steelers home spread -3.5 to +2.5 across 71 books, dispersion 6
- line-c9714ca149113493d7a398fc85aeff83 — Giants @ Red Sox home spread -1.5 across 8 books
- line-4c979daade3eabe626041b803b8ed3dc — Nationals @ Marlins home spread -1.5 across 6 books
- line-970948ad3e20330dce82c2d540759201 — Rays @ Orioles home spread -1.5 across 8 books
- line-8db2519e71d45c54f58d0e3d5ead016f — Panthers @ Jaguars home spread -2.5 to +1.5 across 71 books, dispersion 4
- line-e7241087266735b11507872ed82ce40c — Mets @ White Sox home spread -1.5 across 8 books
- line-a3bba280f85f30544dc7f29ab7717ebc — Tigers @ Royals home spread +1.5 across 8 books
- line-dc562897351ca66ed2c1bca057a99133 — Athletics @ Astros home spread -1.5 across 6 books
- line-d084d47c393be04d4f0ce4668194dce4 — Angels @ Rangers home spread -1.5 across 4 books
- line-b9462d69bbe4a0609c8c894fe1ecd438 — Guardians @ Rockies home spread +1.5 across 8 books
- line-6aeeec35daee963272fbe38448589a76 — Packers @ Broncos home spread -7.5 to -6 across 71 books, dispersion 1.5
- line-7bf7382fa8788015ccab3559e8c13234 — Reds @ Diamondbacks home spread -1.5 across 8 books
- line-42af37af9d3d180bae0e2031201a4a38 — Pirates @ Dodgers home spread -1.5 across 8 books
- line-f4568c9b73d981125c0bdb5283479d57 — Cubs @ Mariners home spread +1.5 across 8 books
- line-d6b25026fff026c4eed65a104032afb4 — Commanders @ Lions home spread -4.5 to -1.5 across 71 books, dispersion 3
- line-e2002d8b700edb5a9b44786e4b1a3b64 — Bills @ Browns home spread -3 across 71 books
- line-3c0361eed8120d457667568f6597016f — Falcons @ Colts home spread -3.5 across 71 books
- line-883b38f5ba3b8125ea9021d432b4cf5f — Ravens @ Vikings home spread -3.5 to -3 across 71 books, dispersion 0.5
- line-da817acdbca0fd0a5537d164587a38aa — Giants @ Dolphins home spread +2.5 to +3.5 across 71 books, dispersion 1
- line-04c3428b8166f2353b0e2d80cef9c286 — Saints @ Rams home spread +1.5 to +3 across 71 books, dispersion 1.5
- line-6dde5642256d4cab5e5a51447d879021 — Bears @ Bengals home spread +1.5 to +2.5 across 71 books, dispersion 1
- line-a2c8b439595e23946ff343bd350673dc — Eagles @ Patriots home spread +1.5 to +2.5 across 71 books, dispersion 1
- line-e967d85a6548f086f54d80e5cefa158a — Chiefs @ Buccaneers home spread -5.5 across 71 books, dispersion 0
- line-ad790c38fc725de6e291eed1e33ea500 — Cowboys @ Cardinals home spread -1.5 across 71 books
- line-ebf859e53b7fa093ec49db48b870f2ac — Seahawks @ Titans home spread -4.5 to -3 across 71 books, dispersion 1.5
- line-d59776a34e9e4929557bdf46d846e0ae — North Carolina @ TCU home spread -8.5 across 38 books
- line-5319a5395f5119ba1227ea2b7b6cfbd0 — North Carolina @ TCU home spread -7.5 across 71 books
- line-672d90c097d14bec14b49dcf2c6c1653 — San Jose State @ USC home spread -38.5 across 71 books
- line-e654a030b2b34bed018f5a896b76cc95 — San Jose State @ USC home spread -38 across 38 books
- line-4ea03e0cf762abff169b83d5dc222308 — NC State @ Virginia home spread -6 to -5.5 across 71 books
- line-0a83af774f4a64cb3d0b295a9a7ff488 — NC State @ Virginia home spread -5.5 across 71 books
- line-b82f3ee7025cbccd287428be9483096d — Jacksonville State @ North Dakota State home spread -7 to -6.5 across 71 books
- line-a9b6e7b705ea849de0327dbe0a0da53e — Jacksonville State @ North Dakota State home spread -7 across 71 books
- line-ad9d926899a1a66bc09e96cf49210fc0 — Sacramento State @ Eastern Michigan home spread -8.5 across 71 books
- line-6fdb7936cba18c7ee6ff6f4a3f21a3c4 — Sacramento State @ Eastern Michigan home spread -9.5 across 71 books
- line-81aa208d45780d18e6b440f1a44951e2 — New Mexico State @ Florida State home spread -31 across 71 books
- line-7283161473d3dc7a5fae1abb31cc8ea7 — New Mexico State @ Florida State home spread -31.5 across 71 books
- line-6bb780ced5423fe0c4e33e8724cd5c38 — Hawaii @ Stanford home spread -6 to -5.5 across 71 books
- line-26ba741caaae6cab4681945800b39e78 — Hawaii @ Stanford home spread -5.5 across 71 books
Bet responsibly — set limits, never chase losses.
Model calibration: predicted vs observed
Predicted win probability bucket vs the empirical win rate inside that bucket on the test set. Points on the y=x reference line are perfectly calibrated; points below mean the model is overconfident in that bucket.
EV per $100 across win rate × odds grid
Expected value of a $100 stake at each combination of true win rate and market odds. Anywhere the cell is positive you have a long-run profitable bet; the magnitude shows how aggressive Kelly will size it.
