Which predicts better with limited data? Comparing learning curve and Bayesian models for improved construction duration estimation

Accurate estimation of construction task duration is crucial for project management. This paper introduces a framework to compare the predictive accuracy of learning curve (LC) analysis and Bayesian estimation under limited on-site information. In a study on high-rise formwork operation, we fitted seventeen LC models across seven unit types, with selection through out-of-sample error on unit values. Bayesian analysis utilized priors that encoded expert judgment error via the coefficient of variation and compared three likelihoods using the Bayes factors. We combined Maximum Likelihood Estimation and Bayesian outputs with the LC in simulation. Analysis was repeated with increasing data, emulating project progress. Prior specification dominated early-stage accuracy, and low expert error amplified gains from well-aligned priors and harms from misspecification, with sensitivity diminishing near 20 data points. Rather than choosing the lowest error LC, it was beneficial to track equations below 10% test error and tighten to 5% with 20 data points. The LC, a point estimator, achieved the lowest continuous ranked probability score. Comparing mean absolute error of probabilistic estimates’ modes also confirmed the LC as the best estimator for this dataset. Introduced framework advances theory and practice through reproducible comparison of productivity models to inform decisions under data scarcity.