Markets gossip constantly. Every second an exchange is open, prices whisper numbers, and almost none of the chatter matters. The useful question — the one this first essay builds toward — is quieter: has the machine generating those numbers recently changed into a different machine? That question is what the rest of the series takes apart, so a shared vocabulary has to come first.

Weather, climate, and the generator

A day’s move is weather. It can be dramatic or forgettable, and it tells you little on its own. A regime is closer to climate: a stretch of weeks or years during which the market’s statistical character stays roughly steady. That character is observable in a handful of numbers:

Typical swing size. How large is a routine day? Volatility has a strong tendency to cluster — turbulent days arrive in bunches, calm days in seasons. Robert Engle’s ARCH model, published in 1982 and later honoured with the Nobel memorial prize, was the first formal capture of exactly this: the market’s variance is itself a process, not a constant.
Directional persistence. Do moves continue or reverse? Trending regimes reward continuation-following behaviour; mean-reverting regimes punish it.
The correlation web. Which assets move together, which offset one another, and when do those relationships hold? A market’s hedging arithmetic lives here.
Response to news. In some regimes, bad news is bought; in others, the same headline starts avalanches. Same information, different physics.

When practitioners say the market “changed regime,” they mean some combination of these shifted far enough, and stayed shifted long enough, that models calibrated on the recent past stopped describing the present.

Three regime breaks worth memorizing

Definitions stay abstract until you attach them to episodes any reader can verify. Here are three well-documented ones, chosen because each broke a different part of the market’s statistical character.

February 2018 — the short-volatility unwind

For years after 2009, selling volatility had behaved like a slow, reliable trade: calm begat calm, and strategies that leaned on that persistence earned steadily for a very long time. On the fifth of February 2018 the calm ended abruptly — the VIX index more than doubled in a single session, its largest one-day close-to-close rise on record, and several exchange-traded products built on persistently falling volatility were terminated within days. Funds whose models had years of quiet in their training windows met a market their assumptions had excluded. The generator had changed; their books had not.

March 2020 — one dash for cash

When the pandemic stress arrived, correlations across assets converged violently. Equities fell; corporate bonds fell; even some safe-haven assets were sold in the scramble for cash in hand. Diversification mathematics — the part of portfolio theory that assumes at least something holds up — stopped operating for weeks before reversing. A regime change is not only higher volatility; it is a change in how losses connect.

2022 — bonds change sides

For roughly two decades, bonds had reliably cushioned equity drawdowns, which is the quiet premise beneath balanced-portfolio arithmetic. As inflation returned and central banks raised rates through 2022, stocks and bonds fell together for most of the year. The negative correlation that a whole style of investing relied on was not a law of nature; it was a feature of a specific regime, and that regime ended.

What exactly “changed”?

Strip the stories away and a detection problem remains. A model sees returns — and only returns — flowing in. A regime change means the probability distribution those returns come from, or the conditional relationships among assets, shifted: the mean, the variance, the tails, the co-movement structure, or the way all of these respond to events. Each detector family you will meet in the next essay picks a different corner of that list and watches it closely.

The working assumption underneath is “piecewise stationarity”: markets are not one stable process, but a sequence of approximately stable processes stitched together by abrupt or slow transitions. Regime detection is the craft of finding the stitching points — using only information that existed at the time.

Why naming regimes is a craft, not a calculation

Here is the trap that gives this series its subject. In a chart viewed years later, regimes look obvious: calm here, chaos there, a clear seam between them. Live, from inside the market, the seam is never labelled. Every detector can only weigh probabilities using the data it has seen so far — and every detector trades speed against false alarms. Faster alarms catch breaks earlier and misfire more; patient alarms misfire less and arrive late.

That trade-off, and the discipline of stating it plainly, is where honest analysis lives. The next essay walks through the four working methods — hidden Markov models, Bayesian changepoint detection, correlation-network monitoring, and distribution-drift statistics — one at a time, with their assumptions visible on the table.