Large language models and machine learning for construction cost accuracy

Purpose The purpose of this study is to introduce an innovative system designed to deliver precise building construction cost prediction before project initiation using historical data. The study addresses the significant cost uncertainties in the UK construction sector, aiming to improve early-stage decision-making, project assessment and cost planning for stakeholders. Design/methodology/approach The study utilised a construction cost dataset to train and validate the prediction models. The methodology began with benchmarking baseline models, then implemented Hybrid Quantum SVR, followed by Hybrid Large Language Models (GPT-2) and XGBoost, concluding in hybrid neural networks, XGBoost and CatBoost. Findings The novel hybrid prediction model achieved outstanding performance metrics, demonstrating near-perfect accuracy and reliability. Key results include a coefficient of determination (R2) of 0.9988, a root mean square error (RMSE) of 812.39 and a mean absolute error (MAE) of 199.66. Significantly, this hybrid approach reduced the RMSE by 90.4% compared to the best-performing base model, confirming its efficacy and superior predictive reliability. Originality/value This research implemented a comprehensive benchmarking strategy that evaluates and contrasts advanced techniques, including Hybrid Quantum SVR and a Hybrid GPT-2 and XGboost model, specifically for construction cost prediction. The study's core originality lies in developing a novel, synergistic machine learning framework that effectively combines the superior regression power of XGBoost and CatBoost with a multi-layer perceptron (MLP). This integrated approach provides a robust, data-driven solution, offering unprecedented predictive accuracy to accelerate the UK construction sector's digital transformation.