Frequency-aware multi-tier machine learning framework for item-level highway construction cost estimation: an improvement on uniform models

Accurate pre-bid cost estimation remains a persistent challenge for state highway agencies, where conventional machine learning approaches apply a single model uniformly across all bid items regardless of their historical data support. This study introduces a three-tier machine learning prediction framework that routes each bid item to a tier-specific algorithm based on historical frequency and unit price prior to model assignment. Training (2017–2022, 20,822 items), validation (2023–2024, 3304 items) and final test (2025, 1492 items from 78 projects) sets were temporally separated. Six tree-based machine learning algorithms (Extra Trees, Random Forest, Gradient Boosting, XGBoost, LightGBM and Decision Tree) and a Hybrid Stacking ensemble were analysed. An exhaustive grid search across 24,696 configurations identified LightGBM for Tier 1 and Extra Trees for Tier 2 as the optimal configuration. At the item level, this tier-based prediction framework achieves a test R2 of 0.897, compared to 0.448 reported by the most recent item-level highway bid price estimation study using multiple linear regression. The prediction framework additionally reduces median absolute error by 39.5% relative to a single-algorithm machine learning baseline of 30.06%. Applied to Wyoming Department of Transportation (WYDOT) pre-bid estimates, the framework supports more reliable funding allocation for highway construction programs.