Answers are grounded in this paper and cite their sources.
Small- and medium-sized enterprises (SMEs) could gain from data-driven pricing, but many have no data scientist on staff, and a reliable model is hard to build without one. A second problem is less visible: a pipeline can look accurate in cross-validation simply because its preprocessing was fitted on the data it is later validated against. In this study, we took a lightweight price-prediction pipeline apart, one step at a time. We used semantic imputation of missing values, hand-built features, a mutual-information filter, cross-fitted target encoding and a blend of XGBoost with Ridge regression. Every data-dependent step was refitted inside repeated nested cross-validation, and the final model was checked on a hold-out set that was set aside before the evaluation, although after the pipeline design was fixed on the full training file. On the Ames housing data (1456 sales), the answer was simpler than we expected. Only the blend mattered: the nested RMSLE fell from 0.1169 (XGBoost) and 0.1134 (Ridge) to 0.1113 (corrected resampled t-test, p = 0.005) and reached 0.1097 on the hold-out set, a median error of about 5% of the sale price. For the other steps, we found no convincing evidence of a change in the accuracy of XGBoost. A full fit ran in under a minute on one CPU core. Because the evidence comes from one dataset, we present the findings as a case study.
No takes yet. Share an insight, caveat, or question.
Popa et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: