elo shap applies Shapley values (fair credit per factor) to an Elo rating (team strength score) so you see each factor's exact point contribution. The market dispersion (book disagreement) tells you whether to trust the breakdown.
What is elo shap and why does it matter?
elo shap breaks a single Elo rating into the exact contribution of every input. A raw Elo number says "Team A is 1520." SHAP says "Team A is 1520 because they beat good opponents (+40), have extra rest (+12), but their QB is out (-35) and they traveled cross-country (-8)." You get a receipt, not a black box. That receipt lets you check the model against the market. The market's dispersion number is your proof.
What the college feed reveals about thin consensus
Our August 23 odds feed covers a full college slate. Almost every game has exactly one reporting book. Arkansas-Pine Bluff at Missouri -54.5 . Bethune-Cookman at UCF -42.5 . Eastern Illinois at Minnesota -43.5 . Long Island at Kansas -38.5 . UTEP at Oklahoma -41.5 . Idaho at Utah -33.5 . Indiana State at Purdue -35.5 . Massachusetts at Rutgers -30.5 . Merrimack at Delaware -28.5 . UAB at Illinois -28.5 . North Carolina A&T at Georgia State -28.5 . Akron at Wake Forest -23.5 . UAlbany at Buffalo -24.5 . West Georgia at Kennesaw State -22.5 . San Jose State at Eastern Michigan -4.5 . Colorado at Georgia Tech -7 . Sacramento State at Eastern Michigan -9.5 . Jacksonville State at North Dakota State -7 . North Carolina at TCU -7.5 . San Jose State at USC -38.5 . New Mexico State at Florida State -31.5 . Hawaii at Stanford -5.5 . Memphis at UNLV -5.5 . NC State at Virginia -5.5 . Seattle at Tennessee -4.5 .
Every one of these shows 1 book and 0 dispersion. In the NFL, 0 dispersion across many books means consensus. In this CFB slice, 0 dispersion across 1 book means we don't know if consensus exists. That distinction changes how you read SHAP.
How to read a SHAP breakdown when dispersion is missing
When dispersion is unavailable (1 book), the SHAP breakdown becomes your primary signal. Look for concentration. If Arkansas-Pine Bluff at Missouri -54.5 puts almost all weight on roster talent differential, that is a single-driver game — the model sees one clear story. If San Jose State at Eastern Michigan -4.5 splits weight across rest, travel, QB status, and recent form with no dominant driver, that is a multi-driver game — the model sees noise.
Our CLV variance study proves single-driver games hit closing-line targets more often. In NFL consensus games (many books, low dispersion), SHAP concentration aligns with market agreement. In CFB thin games, SHAP concentration is your only proxy for conviction. Use it to generate a hypothesis, then verify against injury reports, depth charts, and practice news before you bet. The CFB picks feed logs model CLV against the close so you can track whether the hypothesis held.
When the second book arrives, the dispersion test activates
The same game often appears twice in our feed as more books report. North Carolina at TCU shows -7.5 with 1 book and -8 with 1 book . San Jose State at USC shows -38.5 with 1 book and -38 with 1 book . UTEP at Oklahoma shows -41.5 with 1 book and -40.5 with 1 book . The half-point shift as book count grows is the market processing information. When that happens, dispersion becomes measurable — and you can finally audit whether SHAP's attribution matched the market's move. Learn the Elo math in our NFL betting model guide and the spread mechanics in how NFL spreads work. Track live model performance at the track record page and model notes.
What would change our mind
If a CFB game with 1-book coverage shows SHAP concentrating almost all weight on one factor (Arkansas-Pine Bluff at Missouri -54.5 being purely roster gap) but the line moves 10+ points as more books report, the model overfit the thin data. Conversely, if a game with 1 book shows scattered SHAP but the line never moves when more books appear, the market genuinely had no second opinion — and SHAP correctly reflected that uncertainty. Watch San Jose State at USC (currently -38.5 with 1 book ) — if it holds at -38.5 when 20 books report, the 1-book snapshot was the consensus. If it shifts to -30, the model saw structure the early market missed.
Next time you see a CFB line with one book, pull the SHAP breakdown. If one factor drives most of the rating gap, that's your hypothesis. Verify it against injury reports before you bet.
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.
