Answers are grounded in this paper and cite their sources.
Construction delay prediction is constrained by scarce labeled data, and synthetic records written by domain experts are one proposed remedy. We externally validated one such resource, the CONSTADEX dataset, whose 3530 training records were drawn by a single project manager from normal distributions. Eight learners, including logistic regression, XGBoost, LightGBM, CatBoost, and a stacking ensemble, were trained on the synthetic records using four contract-stage predictors and scored once on 223 field projects. The label rule is undocumented; the field labels matched a delay greater than 20% of the planned duration exactly, and this inferred rule was used. The field ROC AUC ranged from 0.639 to 0.691, with no evidence of a difference between logistic regression and the ensemble (DeLong p = 0.53). Models cross-validated on the field projects reached 0.696, and planned duration alone reached 0.663 because the ratio label places duration in its denominator. Without duration, synthetic-trained models fell to 0.536, and under absolute delay thresholds, the ensemble AUC ranged from 0.483 to 0.605. After recalibration, the ensemble improved the Brier score relative to a constant at the field prevalence by only 0.011, and the decision curves showed uncertain net benefit only for thresholds of 0.70 to 0.80. For this generator, the ratio-defined label created most of the apparent signal, and the synthetic data could not replace local field data.
No takes yet. Share an insight, caveat, or question.
Shamroukh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: