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

Source abstract

Probabilistic machine-learning outputs are increasingly used in financial decision-support systems, yet small or unstable score differences can trigger costly portfolio changes. We develop a two-stage machine-learning decision-support framework for a fixed-size exchange-traded fund (ETF) portfolio. Frozen model scores undergo non-decreasing logistic recalibration on an independent Calibration stage. A zero-slope optimum closes the calibration gate and prohibits discretionary membership replacement; positive-slope groups proceed to Policy Validation, which selects a candidate–holding margin boundary or no active updating. The audit-triggered corrected evaluation covers 18 Chinese equity ETFs, three model classes, and seven non-overlapping evaluation periods. Brier score decreased in 15 of 21 model–period groups, although the mean Monotone-minus-RAW Brier change remained near zero (0.000079); expected calibration error decreased in 17 groups and grouped reliability error in 19. Nine fits retained positive slopes and 12 reached the zero boundary. Among the two prespecified primary models, seven of 14 groups closed at Calibration, five further groups selected no active updating, and two authorized updating. Relative to uncalibrated model probabilities (RAW), monotone paths reduced active replacements from 26 to 11, active-member turnover from 6.50 to 2.75, and mean transaction cost from 25.20 to 19.35 basis points of initial equity. The seven-window economic contrast was positive in two windows, negative in three, and zero in two (mean 0.005187). The framework authorizes or refuses unsupported actions without claiming return enhancement.