Improving Efficiency and Accuracy in Construction Sales Valuation via Random Search Optimization

Abstract

The valuation of construction project sales depends on various economic variables and indices. While accurate cost predictions support financial planning and risk management, traditional grid-search optimization-based machine learning techniques often demand extensive computational resources for training and optimization, especially when large datasets require comprehensive machine learning models. Recent investigations highlighted that random search optimization can shorten the training time of ensemble machine learning methods. Nevertheless, its effectiveness for construction project cost valuation, especially when examining model accuracy and training time, is still unclear. This research examines the usability of random search optimization for machine learning models in construction project sales valuation and compares it with the standard grid search approach. A large dataset with 103 inputs from 372 construction projects is used as the basis of the investigation. Six different machine learning models are designed and optimized under grid search and random search approaches to evaluate training time and predictive accuracy. The study results indicate that random search optimization cuts training time by up to 70% and preserves a high level of accuracy, with the best-performing model achieving an R² of 0.98 on the test set. These findings highlight random search optimization as a strong alternative to grid search, providing significant computational savings without harming model performance. The study offers guidance on effective hyperparameter tuning methods that may facilitate scalable and budget-friendly predictive models for construction project valuation.