Joint Change-Point Monitor

Markets do not announce regime changes one series at a time. A structural break arrives jointly — correlations shift, volatilities re-rank, a dozen channels move together — and a battery of univariate alarms will either miss the joint move or drown you in uncoordinated alerts. The Joint Change-Point Monitor watches the panel as a single multivariate object and, when it flags a break, tells you which channels carried it.

The detection core is working end-to-end today. It runs three complementary detectors — Hotelling-T² statistics with Ledoit-Wolf covariance shrinkage for joint mean shifts, CUSUM for sequential drift, and PELT for retrospective segmentation — over a multivariate panel, and it has produced a hashed internal artifact: a twelve-channel weekly matrix, run through the full pipeline with surrogate-calibrated detection thresholds. The implementation is pure standard library, including hand-rolled, unit-tested linear algebra — no dependency between you and the arithmetic.

What it measures

• Joint breaks. Hotelling-T² over the shrunken covariance detects when the panel’s joint behaviour departs from its estimated regime — the move that no single-series monitor sees because no single series moved far enough alone. Ledoit-Wolf shrinkage keeps the covariance estimate honest when channels are many and history is short.
• Sequential drift. CUSUM accumulates small, persistent departures that a snapshot test would dismiss — the slow bleed into a new regime rather than the clean fracture.
• Retrospective segmentation. PELT partitions the full history into internally consistent segments, so a detected break comes with its context: what regime it ended and what regime followed.
• Per-channel attribution. For every detected changepoint, a decomposition of which series moved — the break attributed channel by channel, not reported as an anonymous multivariate alarm. This is the difference between “something changed” and “rates and the dollar changed; equities did not”.
• Calibration by surrogate. Detection thresholds are calibrated against surrogate data rather than asymptotic formulas, so the false-alarm behaviour is measured, not assumed. The calibrated thresholds and the detected changepoints on published panels: NOT YET PUBLISHED — they ship with the hashed artifacts that prove them.

What it does not do

It does not predict the next break. Change-point detection is the honest accounting of breaks that have occurred — sequentially, as early as the evidence permits, but never before it. Any instrument that claims to detect a regime change before the data contains one is describing its own noise.

Who it is for

Portfolio and risk managers monitoring many series at once, for whom the dangerous move is the coordinated one; anyone who has learned that twelve univariate alarms are not one multivariate answer.