Monotone Probability Recalibration for Turnover-Constrained ETF Portfolio Updating: A Two-Stage Machine Learning Decision Support Framework

Machine learning models can rank exchange-traded funds (ETFs) by predicted future performance, but a higher rank does not establish whether replacing a current holding is worthwhile after trading costs. We study a two-stage replacement rule using the predicted probability of an ETF entering the top 30% by its next ten-session return. First, a logistic recalibration constrained not to reverse the ETF ranking adjusts these probabilities; if it assigns every ETF the same probability, replacements are prohibited. Second, a separate policy validation period selects a minimum candidate–holding probability difference for trading or selects no active updating. The analysis covers 18 surviving Chinese equity ETFs, three models, and seven evaluation windows. Earlier account results could not be reproduced from the retained verified inputs. We therefore report results from a separate reconstruction. Because this reconstruction was specified after the earlier evaluation outcomes had been inspected, the analysis is retrospective and exploratory. Across 14 primary model-by-window combinations, the mean Brier loss decreased from 0.227 to 0.225 and the ten-bin expected calibration error from 0.051 to 0.031, without a uniform improvement. Five groups prohibited replacements at calibration, two more selected no active updating, and seven selected finite trading thresholds. The proposed policy made 47 replacements versus 39 using uncalibrated probabilities with the same validation procedure; the mean annualized certainty-equivalent returns were 0.019 versus 0.031. No-update and fixed-threshold comparisons also provide no general economic superiority evidence. The framework makes the conditions for acting on a forecast explicit, while its investment value requires prospective evaluation on a broader historical universe.