Prediction of Construction Material Prices Using Recurrent Neural Networks and Ensemble Batch Intervals

Shifting global trade dynamics and the consequent supply chain bottlenecks complicate the accurate estimation of construction material prices. The goal of this paper is to provide reliable predictions of material prices. Although studies have introduced models for forecasting material prices, the stringent assumptions regarding residual properties pose challenges to their robustness. To this end, the authors adopted a multi-step methodology that included developing multivariate long short-term memory (LSTM) models to predict the prices and confidence intervals on the forecasted prices using a distribution-free ensemble batch prediction algorithm. Results indicate that the LSTM models can forecast the price trends of each of the cement, steel, and lumber prices in the US with a mean absolute percentage error of less than 4%. The ensemble batch confidence intervals provide a range that contains the actual prices in the forecasting horizon. The proposed approach helps the stakeholders in making better-informed predictions.