Source Exclusion and Regime-Gated Equity-Factor Forecasts

Summary

The audit examines how to remove a predictor from regime-gated forecasts. Deleting a variable’s lagged inputs is not equivalent to excluding it from the full fitting, filtering, and tuning process, because the variable can also affect the estimated regime gate. The researchers reconstruct an HML-to-SMB forecast relation and expand the analysis across 30 factor directions in four regional panels, using pooled and hidden Markov model forecasts. They also compare shared-gate ablation with complete source exclusion and use simulations to distinguish underlying predictive information from costs introduced by fitting the model.

The reported evidence is mixed and generally inconclusive: the reconstructed relation has negative incremental mean-squared-error gains, while ablation and full exclusion yield opposite capped-loss point-estimate signs in 42 of 240 comparisons. None of 2,400 temporal endpoints resolves a direction under the simultaneous confidence procedure. Simulations show fitted HMM losses even when the oracle signal is beneficial. Correlated markets, retrospective data vintages, limited inferential resolution, and unverified external exposure constrain confirmation claims.

Key ideas

• A source can affect both forecast coefficients and the estimated regime gate, so removing only its lagged inputs is incomplete exclusion.
• The audit compares shared-gate ablation with exclusion throughout fitting, filtering, and tuning.
• The reported financial comparisons produce conflicting point-estimate directions and no resolved temporal endpoint.
• Simulations show that fitting costs can outweigh positive predictive information available to an oracle model.
• Correlated markets and retrospective data vintages limit the strength of confirmation claims.

Tags

# Source Exclusion in Regime-Gated Forecasting: A Cross-Market Audit of Equity-Factor Predictability


# Source Exclusion in Regime-Gated Forecasting: A Cross-Market Audit of Equity-Factor Predictability









A source variable can enter a forecast through both lag coefficients and an estimated regime gate. Removing its lag block therefore differs from excluding its information throughout fitting, filtering and tuning. We audit this distinction in equity-factor forecasting. A current-vintage reconstruction of the reported HML-to-SMB relation finds negative incremental MSE gains for pooled, Gaussian-HMM and Student-t-HMM forecasts in 2010-2024. An expansion evaluates all 30 factor directions in four regional panels, with 2025-August 2026 withheld from earlier repository forecast evaluation. Shared-gate ablation and complete exclusion give opposite capped-loss point-estimate signs in 42 of 240 unweighted temporal comparisons. None of 2,400 temporal endpoints resolves a direction under a simultaneous confidence-sequence construction. All 776 intervals within a declared score-resolution margin are explained by forecast proximity. Controlled simulations separate oracle information from fitting costs and show fitted HMM losses despite positive oracle information. A matched Gaussian-versus-Bernoulli channel intervention improves forecasts at one corruption level under known-AR residualization, without identifying the earlier pipeline failure or a financial mechanism. Application-budget and learned-pipeline diagnostics expose poor resolution of known positive targets. Independent implementations replay archived forecasts and recompute all 4,800 primary financial interval summaries. The evidence demonstrates sensitivity to source removal and limited inferential resolution. Correlated markets, retrospective data vintages and unverified outside exposure limit confirmation claims.

Shown in full with attribution under the source's licence. Licence: abstract CC0

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.