Construction cost prediction using TabPFN-based stacking model with SHAP analysis

Purpose Construction cost prediction is essential for improving budget management, risk reduction, and resource allocation in the construction industry. This study aims to enhance prediction accuracy by proposing a stacked ensemble model that integrates Tabular Prior-data Fitted Network (TabPFN), CatBoost, and Gradient Boosted Decision Trees “GBDT” while incorporating SHapley Additive exPlanations (SHAP) value analysis for interpretability. Design/methodology/approach This study develops a two-layer stacking ensemble model using San Francisco building permit data, where Estimated Cost denotes the permit-declared pre-commencement cost estimate. TabPFN and CatBoost are used as first-layer base learners, and their out-of-fold predictions are combined with original project features as inputs to a second-layer GBDT meta-learner. Model performance is evaluated using 10 independent random train-test splits. In each run, the training set is used for leakage-controlled 5-fold cross-validation and model tuning, and final performance is reported as the mean ± standard deviation across the 10 runs. RMSE, MAE, and MAPE were employed in comparison with baseline models, and SHAP values are applied for model interpretation. Findings The proposed TabPFN + CatBoost-GBDT model achieves robust performance, with a MAPE of 1.89% for recorded permit-estimated costs. This indicates close alignment between predicted values and declared pre-commencement cost estimates, although the results should not be interpreted as final realized construction costs. Additional diagnostics show that high-cost projects remain more difficult to predict. The low prediction error indicates that the model can provide more reliable pre-commencement cost estimates, which may assist contractors in bid pricing, budget control, contingency planning, and resource allocation by reducing uncertainty in expected project expenditures. Originality/value This study extends pretrained tabular learning to construction cost prediction by integrating TabPFN, CatBoost, and GBDT in a task-specific stacking framework. It combines pretrained tabular representation, categorical feature handling, and feature-informed meta-learning. SHAP analysis is further used to explain how the final model combines first-layer predictions and original project variables, while raw feature interpretation is treated as conditional and partly indirect.