Freight pricing in brokerage platforms is a challenging regression problem characterised by high heterogeneity across shipment types, temporal non-stationarity of costs, and sparse multi-quotation structures. We present an end-to-end machine learning pipeline that addresses these challenges through four methodological contributions: (i) a domain-driven price actualisation formula that decomposes inflation into fuel and macroeconomic components; (ii) a similarity-based matching algorithm that groups single-quotation orders with comparable shipments, enriching the effective training signal; (iii) a stratified modelling strategy that trains separate Random Forest regressors for three structurally distinct shipment categories—Full Truck Load, Less-than-Truck Load, and Groupage; and (iv) a Gaussian Mixture Model-based confidence estimator that quantifies prediction reliability. Evaluated on a real-world dataset of approximately 58, 000 historical quotations spanning nine years, our system achieves a Mean Absolute Percentage Error of 3.23% for full loads, 5.60% for partial loads, and 10.76% for groupage shipments on a held-out temporal test set, representing a 57–68% relative improvement over the incumbent rule-based pricing engine.
A Machine Learning Pipeline for Freight Transport Pricing / Bottino, A., Cuomo, S., Vocca, V.. - 3019:(2026), pp. 423-438. (21st International Conference on Information Processing and Management of Uncertainty in Knowledge-Based Systems, IPMU 2026 ita 2026) [10.1007/978-3-032-28994-0_30].
A Machine Learning Pipeline for Freight Transport Pricing
Bottino A.;Cuomo S.
;Vocca V.
2026
Abstract
Freight pricing in brokerage platforms is a challenging regression problem characterised by high heterogeneity across shipment types, temporal non-stationarity of costs, and sparse multi-quotation structures. We present an end-to-end machine learning pipeline that addresses these challenges through four methodological contributions: (i) a domain-driven price actualisation formula that decomposes inflation into fuel and macroeconomic components; (ii) a similarity-based matching algorithm that groups single-quotation orders with comparable shipments, enriching the effective training signal; (iii) a stratified modelling strategy that trains separate Random Forest regressors for three structurally distinct shipment categories—Full Truck Load, Less-than-Truck Load, and Groupage; and (iv) a Gaussian Mixture Model-based confidence estimator that quantifies prediction reliability. Evaluated on a real-world dataset of approximately 58, 000 historical quotations spanning nine years, our system achieves a Mean Absolute Percentage Error of 3.23% for full loads, 5.60% for partial loads, and 10.76% for groupage shipments on a held-out temporal test set, representing a 57–68% relative improvement over the incumbent rule-based pricing engine.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


