Machine learning study demonstrates improved pre-bid cost estimation accuracy in highway construction projects, suggesting more reliable state infrastructure funding allocation.
Key Points
To develop and evaluate a frequency-aware, multi-tier machine learning framework that assigns highway bid items to tier-specific predictive models based on historical frequency and unit price.
Evaluated six tree-based algorithms (Extra Trees, Random Forest, Gradient Boosting, XGBoost, LightGBM, and Decision Tree) alongside a hybrid stacking ensemble via grid search across 24,696 configurations.
Trained and tested models on temporally separated records from the Wyoming Department of Transportation, using 20,822 items (2017–2022) for training, 3,304 items (2023–2024) for validation, and 1,492 items from 78 projects (2025) for testing.
Identified LightGBM (Tier 1) and Extra Trees (Tier 2) as the optimal setup, achieving a test R2 of 0.897 relative to 0.448 for standard multiple linear regression.
Reduced median absolute prediction error by 39.5% compared to a single-algorithm machine learning baseline of 30.06%.