Regime Classifier — Changelog

Why we publish a changelog

A regime classifier is a statement about hidden state. When the model says “Bull regime, 78% confidence,” the reader is taking on faith that the state space, the training window, and the inference procedure have not silently changed underneath them between yesterday’s call and today’s. A probabilistic three-state regime classifier trained on a long historical window is a different object from a smaller-state model trained on a shorter window, even if both label today as “Bull.”

This page is the public record of every methodology change to the AZTMM regime classifier. It is append-only. If the state space changes, the training window changes, the feature inputs change, the initialization procedure changes, or the smoothing rule changes — it shows up here. Routine daily inference does not.

How to read this

Entries are dated descending. Each entry is scoped to a methodology change: state space, training window, feature set, training procedure, or visualization changes that affect interpretation. Daily regime probability updates and routine state transitions are not logged here — those live in the Pulse posts.

The log

2026-05-08

Regime classifier methodology v2.0 published. Probabilistic three-state regime model with standard regime-fitting techniques and posterior confidence calculation documented.

2026-05-07

Daily refresh schedule formalized at 4:15 PM ET (post-close). Regime-persistence framework introduced.

2026-04-15 → 2026-04-21

Phase 5 audit identified Pulse Lab regression on regime persistence section; redesigned visual model and confidence chart.

2026-04-10

Training window standardized at a long historical window of broad-market pricing data. Multiple random initializations.

2026-03-20

Three-state regime classifier (Bull / Sideways / Bear) committed. Switched from two-state experimental to three-state production.

2026-02-15

Regime visualization prototype shipped (state probabilities + persistence-framework display).

Future scheduled changes

• Q3 2026 — expanded feature set (evaluation). Currently the regime classifier is trained on broad-market pricing data. We are evaluating whether adding credit spreads (HYG/LQD) and yield curve slope as conditioning variables improves transition-regime detection.
• Q4 2026 — walk-forward validation publication. Once the regime classifier v2.0 has been live for 6+ months, walk-forward results will be published on the Performance Archive.
• Q4 2026 — lag-during-fast-regime-change documentation. The Jan 2025 whipsaw episode is documented as a known model failure mode. Q4 2026 will publish a quantified analysis of classifier lag during fast regime transitions.