Probabilistic forecasting of construction labor productivity metrics

This study investigates the possibility of doing probabilistic forecasting of construction labor productivity metrics for both long-term and short_term estimates. The research aims to evaluate autoregressive forecasting models, which may help decision-makers with information currently unavailable in construction projects. Unlike point forecasts, the proposed method employs probabilistic forecasting, offering additional valuable insights for decision-makers. The distributional information is obtained by updating the moments of the distribution during training. Two datasets are used to evaluate the models: one collected from an entire construction site for long-term forecasting and one from an individual worker for short_term forecasting. The models aim to predict the state of direct work, indirect work, and waste. Several models are trained using different hyperparameters. The models are tuned on the number of trees and the regularization used. The presented method gives estimates of future levels of direct work, indirect work, and waste, which will add value to future processes.