Elo and SHAP answer different questions. Elo summarizes team strength from results. SHAP explains how the features inside a particular fitted model divided responsibility for a particular prediction.
Neither tool makes a model honest by itself. Elo can be stale or underspecified. SHAP can explain a bad model perfectly. The useful pair is a rating, an attribution, and a public grading rule that checks whether the predictions work.
Elo is a model family, not one sacred number
An Elo system carries a rating from one game to the next. The update depends on what the system expected and what happened. An upset usually moves the state more than an expected result. The exact rule belongs to the model contract.
Models can choose different starting values, update sensitivity, margin treatment, home adjustment, season carryover, and neutral-site handling. Two Elo models can watch the same games and disagree without either violating the concept.
The rating is retrospective. It does not automatically know about an injury, weather, travel, or a coaching change unless the broader prediction model supplies those inputs separately.
SHAP explains the fitted model relative to a baseline
SHAP uses Shapley-value logic to allocate the difference between a model’s baseline output and a specific prediction across the features the model received. Positive attribution pushes the output one way; negative attribution pushes it the other.
That is an explanation of model behavior, not a causal estimate of the real world. If two features carry similar information, the attribution can be shared or shifted depending on the explainer and background data.
A clean waterfall chart therefore answers “what moved this model output?” It does not answer “what caused the game result?” or “what did the market fail to price?”
The market rows are not SHAP reports
The attached Chiefs–Buccaneers row and Jets–Steelers row show that historical game-market observations exist. They do not show which features an Elo model used or what attribution any feature received.
The old copy looked at the market and reverse-engineered a confident story about quarterbacks, rest, travel, and roster quality. That is narrative, not explainability. A source row proves the quote. A model schema proves the inputs. A stored SHAP artifact proves the attribution.
Keep those layers separate. Otherwise the explanation becomes a postgame story wearing a bar chart.
Transparency does not validate the model
Suppose SHAP shows that one feature dominated a prediction. That means the feature moved the fitted output under the explainer’s assumptions. It does not mean the feature should have moved it, that the data was available pregame, or that the market missed it.
Validation needs frozen predictions and an evaluation contract. Over a named window, report ATS wins, losses, win rate, and n when those records exist. A batch stamp is not a time axis, and a single game is not a record.
Calibration and out-of-sample grading judge the forecast. SHAP diagnoses the model. Those jobs complement each other precisely because they are not the same.
Feature provenance is part of the explanation
A pregame model should use an allow-list that excludes information created after kickoff or settlement. SHAP will faithfully explain leaked data if the pipeline feeds it leaked data.
Every attribution should travel with feature name, source, cutoff time, transformation, baseline, model version, and explainer settings. Without that context, the receipt cannot be reproduced.
The glass-box guide explains why inspectability matters. The closing-line guide supplies one market benchmark after the prediction is frozen.
The failure rule belongs in the contract
Before publishing attributions, decide what would reduce trust: calibration drift, unstable feature signs across adjacent windows, provenance violations, or a graded record that misses its stated threshold over the required n.
Those rules must be written before results arrive. Otherwise the explanation layer becomes a machine for justifying whichever output the model produced.
When the stored model artifact is missing, return EXPLANATION UNAVAILABLE with a cure: persist the model version, baseline, features, and attribution output for the frozen prediction.
What the pair is genuinely good for
Elo gives a compact, updating view of team strength. SHAP gives an analyst a way to inspect why a fitted model departed from its baseline on a specific prediction. Together they make a forecast easier to challenge.
That challenge should be concrete. Was the rating current? Were the inputs available pregame? Did correlated features split attribution strangely? Did the model perform over a named sample?
The market rows can provide comparison context. They cannot prove which model reason was correct. Only stored predictions and settled grading can tell us whether the explanation belonged to a useful model or merely a transparent one.
