Both curves come from the same fitted slope — only the intercept differs. The dashed red curve is what the model reports straight out of training on the balanced sample: at an average score (0), it predicts the sample's balanced rate ȳ, not the real-world rate. The solid green curve applies the King & Zeng (2001) prior correction, shifting the intercept by ln(τ/(1-τ)) − ln(ȳ/(1-ȳ)) so that an average-scoring case is predicted at the true population prevalence τ instead.

Drag ȳ down toward τ (less aggressive balancing) and the two curves converge — the correction shrinks because there's less artificial imbalance to undo. Push τ toward a rare-event regime (fraud, cancer, default) with ȳ still at 50/50, and the gap between the curves — and the risk of shipping an uncorrected, badly over-predicting model — grows fast.