DOI RECORD
Graph-Aware Rail-Freight Forecasting, Empirical Uncertainty Calibration, and Distributionally Robust Investment Prioritization: A Finland–Kazakhstan Study
Abstract
Rail-freight infrastructure planning requires decisions under predictive uncertainty and network interaction. This study develops a country-adaptable framework linking graph-aware freight forecasting, chronological network-level conformal calibration, capacity-constrained railway response, economic appraisal, and Wasserstein distributionally robust portfolio optimization. The predictive layer uses 56 monthly Fintraffic/Digitraffic archives from January 2022 to August 2026, containing 464,823 Cargo train records. The chronologicalprotocol assigns 2022–2023 to topology selection and model training, 2024 to uncertainty calibration, 2025 to model and ambiguity-radius validation, and January–August 2026 to held-out primary evaluation. Graph-aware Ridge achieved a held-out MAE of 1.05055 trains/day, 0.549% below the validation-selected non-graph Ridge. A 7-day moving-block bootstrap gave a 95% interval of [0.000199, 0.011456] trains/day for the MAE gain. The network conformal set attained 97.94% empirical simultaneous coverage for a 90% target; the source remains unresolved. The investment layer evaluates all 16 portfolios of four Kazakhstan rail projects. Removing the P2-based calibration of the avoided/deferred-demand value and varying μU from USD 5 to USD 15 changes the exact optimal portfolio and decision boundaries, but P2 remains included in every optimum over the resulting 79,380-cell calibration–ambiguity–risk grid. Infrastructure ranking therefore depends on uncertainty, network constraints, and calibration assumptions, while P2 forms the robust investment core over the examined family.