Caltrans Granular Forecasting v1.0.0: Reproducibility Package for Leakage-Safe Highway Award-Occurrence Ranking and Conditional Unit-Price Forecasting

Abstract

This record contains the reproducibility package for the study “A Leakage-Safe Two-Stage Decision-Support Framework for Granular Highway Award-Occurrence Ranking and Conditional Unit-Price Forecasting.”The study develops a leakage-safe two-stage decision-support framework for granular highway procurement forecasting using California Department of Transportation (Caltrans) awarded-bid records. The primary stage ranks monthly material–item–standardized-unit–district series according to the likelihood of an award, while the second stage provides conditional unit-price estimates when an award is expected.The analysis covers audited Caltrans awarded-bid records from 2011 through May 2026 across all 12 Caltrans districts and ten highway-construction material groups. The framework incorporates calibrated known-series occurrence ranking, a separate cold-start district-diffusion pathway, conditional unit-price modeling, quantity-aware price analysis, hierarchical conformal uncertainty, explainability, feature-group ablation, operational review-queue evaluation, statistical comparison, and scale-aware diagnostics.The archive contains the recovered core Python analysis scripts, exact archived explainability output tables, deterministic publication figures, compact numerical tables supporting results reported in the manuscript and supplementary material, reproducibility documentation, licensing information, and SHA-256 integrity manifests.Raw Caltrans awarded-bid records are not duplicated in this archive because they are publicly available through the California Department of Transportation Contract Cost Data portal. The included acquisition, audit, preprocessing, and modeling scripts document the workflow used to reconstruct the analysis from the public source, subject to future availability and consistency of the Caltrans portal.The frozen January–May 2026 evaluation produced an average precision of 0.4557 and ROC-AUC of 0.9455 for the known-series occurrence model. Reviewing approximately the top 10% of known-series candidates captured 73.57% of known-series award events, 80.85% of corresponding awarded value, and 87.51% of corresponding awarded quantity. Unknown-quantity conditional-price forecasting achieved 56.23% sMAPE, while the quantity-known secondary scenario achieved 40.99% sMAPE.Important scope: the occurrence model is intended primarily for procurement candidate prioritization rather than deterministic prediction of individual awards. Conditional unit-price estimates are intended as broad decision-support guidance rather than precise project budgets. Explainability measures describe model-specific predictive associations and should not be interpreted causally.Some later one-off analyses were not preserved as standalone executable scripts in the recovered project state. For those analyses, this archive preserves the exact numerical output tables reported in the manuscript rather than claiming an executable rerun that is not available.