Abstract
Text-based indices from news and social media track latent constructs such as risk, sentiment, and uncertainty, but evolving language and coverage induce temporal measurement error that undermines longitudinal comparability.We cast calibration as a rank-constrained convex program: with partial ordinal information on a subset of time points, a convex monotone M-estimator yields a unique rank-consistent fit; the squared-loss specialization reduces to pool-adjacent-violators estimation with standard isotonic guarantees.Stochastic residual augmentation preserves expected ordering while restoring variability, and a regularized supervised model propagates the calibrated scale to the full series without additional rank labels.A semiconductor risk case study with expert anchors shows strong leave-one-out generalization, alignment with independent volatility proxies, and improved temporal coherence relative to distribution matching and regression-style comparators.

This publication has 20 references indexed in Scilit: