A SHAP waterfall plot for one loan applicant shows: base value 0.12 (the average default probability), then contributions +0.21 from debt_to_income, +0.06 from prior_delinquency, −0.09 from years_employed, and −0.02 from age, ending at a prediction of 0.28.
How should a loan officer read this plot?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
A waterfall plot is a local explanation: it decomposes one prediction into feature contributions that sum from the average (0.12) to this applicant's score (0.28 = 0.12 + 0.21 + 0.06 − 0.09 − 0.02). It answers "why this score for this person?" — exactly what an adverse-action notice or a customer conversation needs. It is not a ranking across the portfolio and not a causal what-if.
That is a claim about the whole population, which needs the beeswarm or mean |SHAP| plot. For a different applicant a different feature could dominate.
SHAP attributes the model's output; it does not predict what happens if the input changes. Correlated features make the two differ.
0.28 is the predicted probability of default, not a measure of how correct the model is.