What question did this study set out to answer?

The aim is to enhance construction cost prediction accuracy while incorporating uncertainty and interpretability in the model.

March 29, 2026Open Access

Uncertainty aware and explainable construction cost prediction using a hybrid probabilistic learning model

Key Points

The aim is to enhance construction cost prediction accuracy while incorporating uncertainty and interpretability in the model.
Developed a unified probabilistic framework called NGBoost-ETR.
Trained on a real-world dataset of 4477 samples from RSMeans.
Evaluated using various metrics for predictive calibration and uncertainty estimation.
Achieved R^2 = 0.9866, demonstrating high predictive performance.
Validated superior uncertainty calibration across multiple quantitative metrics.
Outperformed baseline models and other NGBoost-based hybrids in interval efficiency and overall calibration.

Abstract

This paper presents a unified probabilistic framework for construction cost forecasting, NGBoost-ETR (Natural Gradient Boosting with Extra Trees base learners) that delivers predictive accuracy, calibrated uncertainty, and SHAP-based interpretability. Rather than positioning novelty in algorithmic integration alone, the contribution lies in a systems-level design that jointly addresses three critical gaps in cost modeling: reliable interval calibration, model interpretability, and robustness across complex feature interactions in domain such as sustainable building. Trained on a real-world RSMeans dataset of 4477 samples, NGBoost-ETR achieves superior predictive performance (R2 = 0.9866, RMSE = 0.4986, MSE = 0.2486, MAE = 0.2300, and MAPE = 1.4314%) compared to 10 baseline regressors and 9 NGBoost-based hybrids. Beyond point prediction, the model also demonstrates robust probabilistic calibration, validated through a comprehensive suite of six quantitative metrics. Specifically, evaluation based on Prediction Interval Coverage Probability (PICP), Prediction Interval Normalized Average Width (PINAW), Mean Prediction Interval Width (MPIW), Coverage Width-based Criterion (CWC), Negative Log-Likelihood (NLL), and Continuous Ranked Probability Score (CRPS). In head-to-head ablations, NGBoost-ETR attains the best interval efficiency (lowest PINAW, MPIW), overall calibration (lowest CWC), and distributional accuracy (lowest CRPS), with competitive NLL and acceptable coverage-outperforming variants that inflate coverage via impractically wide intervals. Crucially, not all hybrids are beneficial (e.g., some NGBoost pairings underperform their base learners), emphasizing that the ETR pairing is a validated choice rather than a generic integration. This work introduces a process innovation by embedding reliable uncertainty estimation into tree-based models without sacrificing performance, supporting greater resource efficiency in cost planning and estimation. The resulting framework not only supports data-driven budgeting and tendering but also promotes transparent institutions and risk-aware decision-making in construction management.

AI에게 질문

Bookmark

View Full Paper