Regime Survival Forecasting for Adaptive Execution: Beyond FixedAggregation Windows

Abstract

We test whether Weibull Accelerated Failure Time (AFT) survival models can replace the fixed aggregation
window in Hidden Markov Model (HMM)-based regime-aware execution with a per-instance prediction
of remaining bearish regime duration. Fitting models to 54 bearish regime instances across eight asset
classes (2020–2024), we find the extension fails on three structural grounds. First, concordance indices of
0.20–0.39 confirm that HMM-derived covariates, posterior entropy and stay probability at regime start,
carry no duration-predictive information under parametric AFT models. Second, a without-replacement
subsampling simulation shows the concordance index remains flat at approximately 0.48 from n = 4
to n = 45 training instances, ruling out data quantity as the cause. Third, Weibull shape parameters
below 1.0 in every asset class produce decreasing-hazard distributions whose mean durations structurally
exceed the window cap, collapsing 60–89% of predictions to boundary values regardless of covariate values.
There are no genuine adaptive wins over the fixed ten-day aggregation window established in prior work.
Together, Papers I-III characterize the limits of HMM-based regime awareness in execution: regime
signals require multi-day aggregation (Paper II), learned policies cannot exploit them reliably (Paper I),
and adaptive window calibration via parametric survival models is not achievable with HMM posteriors
as duration predictors (Paper III). These findings close a research trilogy with a complete empirical
characterization of what HMM-based regime awareness can and cannot achieve in optimal trade execution.