Forecasting construction costs with small datasets: a decision-making framework with ensemble learning approach

Purpose Construction cost forecasting is vital for early-stage project decisions, including feasibility analysis and bid preparation. This study addresses the challenge of achieving accurate cost predictions using small datasets – an enduring limitation in construction data practices. Design/methodology/approach The research proposes a decision-making framework that applies a bagging-based ensemble learning (EL) approach using Support Vector Regression (SVR) and Artificial Neural Networks (ANN) as base learners. The framework is validated across four real-world construction datasets, each with fewer than 150 data points. Findings The ensemble model significantly outperformed single learner models in terms of prediction accuracy. For instance, the Mean Absolute Percentage Error (MAPE) in Dataset 1 decreased from 46.32% (SVR) and 17.84% (ANN) in individual models to 17.84% and 12.92%, respectively, with EL. Similarly, Dataset 2 saw a reduction from 36.62% to 11.41% (SVR) and from 18.59% to 11.41% (ANN). Research limitations/implications While the proposed method is highly effective for small datasets, its computational efficiency and accuracy on large datasets warrant further exploration. Practical implications The framework provides a structured, data-driven approach for practitioners and cost estimators to improve early cost predictions when faced with limited data. Originality/value This study is among the first to develop and validate a decision-making framework for construction cost forecasting specifically tailored for small datasets. It also quantifies the effect of varying ensemble and bag sizes – an area previously underexplored in construction cost forecasting literature.