Horizon-dependent pavement roughness forecasting with machine learning and treatment history

Predictions of pavement roughness support maintenance planning, but apparent model skill can be overstated when repeated section observations, persistence and intervention chronology are ignored. Fixed one-, two- and three-calendar-year International Roughness Index (IRI) targets were reconstructed from Long-Term Pavement Performance Standard Data Release 39 and linked to official maintenance and rehabilitation records. The primary sample excluded intervals with possible treatment crossings, and treatment covariates were restricted to records before the forecast origin. Five nested rolling-origin evaluations compared persistence with CatBoost using condition, structure and materials (P2) and the same variables plus treatment history (P2T). P2 did not improve on persistence at one or two years, but reduced MAE by 0.0104 m/km at three years. P2T did not improve P2. Section-maximum 90% conformal coverage for P2 was 0.942, 0.924 and 0.902. All rapid-deterioration cases were underpredicted. Machine-learning value depended on forecast horizon.