Estimation of Residential Building Repair Costs Using Selected Machine Learning Algorithms

This study examines the feasibility of predicting net repair costs (Estimated Cost, PLN) for multi-family residential buildings from data extracted from technical inspection reports. Rather than merely comparing algorithmic performance, the analysis was designed as a diagnostic sequence aimed at identifying the sources of prediction error. Four machine learning algorithms (Extra Trees, Random Forest, XGBoost, and GBM) were first applied to direct regression of repair cost. We then tested whether the difficulty of estimating exact cost values stems from the high variability of the target variable and whether this limitation can be mitigated by a two-stage approach: assigning observations to one of three cost-risk bands (Low, Moderate, High) and subsequently estimating cost within the assigned band. The empirical cost distribution was strongly right-skewed (median: 4500 PLN; mean: 63,402 PLN; maximum: 3,680,524 PLN). The best direct regression model achieved an R2 of 0.452, while the best fully deployable two-stage model, combining an XGBoost classifier with a Random Forest regressor, achieved an R2 of 0.392. When it was assumed that the actual cost-risk bands were known, an R2 value of 0.839 was obtained, indicating that the main source of error is not regression within the bands, but rather the initial stage of assigning the bands. These results demonstrate that reporting a single global R2 for highly skewed, weakly identifiable cost data can be misleading, and that decomposing predictive performance into band-assignment and within-band regression components provides a more informative evaluation. This article also points to concrete directions for improving the underlying database, particularly through the inclusion of variables describing repair quantity, unit of measure, and detailed repair scope.