Independent editorial · AI in investment
Drift Investment Insights
A free essay series about one narrow craft: how machine-learning methods detect market regime changes — the quiet moments when markets change character.
What changes when a market changes regime
The subject the essays keep circling: a market whose underlying statistics shift beneath everyone's feet.
Short-term moves are weather; a regime is climate. In quantitative terms, a regime is the statistical character of a market — the typical size of daily swings, the way stretches of calm and stress follow one another, the pattern of what moves together — sustained long enough that models and people start treating it as the background condition.
Regime change is the moment that background ends. In February 2018, funds that systematically sold volatility had enjoyed years of calm — until the day the VIX index, in its largest one-session rise on record, ended at more than twice its prior close, and their models' learned world stopped describing reality. In March 2020, nearly every asset class fell together in a worldwide dash for cash. Across 2022, bonds stopped cushioning equity losses as inflation returned, and two decades of portfolio arithmetic quietly broke. In each case the process behind prices changed first; the headlines caught up later.
That gap — between what a market already is and what models still assume — is what regime-change detection exists to close, and what this series keeps returning to.
Four detector families the essays return to
Regime detection is not one technique but a small guild of them. Each method below gets honest treatment in the library — including the limit the desk insists on stating.
Hidden Markov models
The market is modelled as hopping among hidden states — calm, trending, stressed — and every close updates the probability of which state is active. Hamilton's 1989 switching model started this line of work.
The limit: states that models find are clusters; people do the naming. "Crisis" is a label, not a discovery.
Bayesian changepoint detection
After each new observation the model recomputes the odds that the process generating returns just shifted — the run-length formulation of Adams and MacKay (2007) is the classic reference.
The limit: alarms tuned for speed are also quick to false-fire on noisy stretches.
Correlation-network monitoring
Represent markets as a web of what-moves-with-what and watch for ruptures in the web — hedges quietly failing is often the earliest symptom of a new regime settling in.
The limit: edges measured on short windows are noisy threads that snap by accident.
Out-of-distribution drift
Statistics like the population stability index compare today's distribution of returns or factor exposures with the window a model was trained on — a staleness alarm for any investment AI in production.
The limit: drift is only confirmed after the new reality has already accumulated.
Who this desk is writing for
A reading room works when everyone knows what it is — and is not — for.
Readers who enjoy explanations one honest level deeper than the headlines: analysts refreshing quantitative fundamentals, engineers building models who want the investment meaning, students, and the genuinely curious. The essays assume no mathematics beyond comfort with a chart and a sentence of statistics.
It is not written for anyone hunting personalized investment advice, trading signals, or managed money. Nothing here recommends an asset, a timing, or an account; nothing on the site carries a price — no subscription, no course, no deposit, no managed portfolio.
Start with the primer
Ask the inquiry desk
Questions about a method, corrections, a topic the series should take apart next — all welcome.
The inquiry desk handles correspondence only. Ask about the assumptions behind a detector, flag an error, argue with a framing, or propose the next essay subject.
Sending an inquiry costs nothing, and nothing is ever solicited back — there are no services, products, portfolios or paid tiers behind it.
Fair questions, brief answers
The objections readers raise most, answered plainly.
Is any of this investment advice?
No. The essays explain published methods, public research and market history. Nothing here is a recommendation or an offer, and no essay is personalized to any reader's situation.
Does anything on this site cost money?
No. Reading is free, inquiries are free to send, and the desk never solicits payment, subscription or deposit of any kind — there is nothing to buy and no checkout anywhere.
What happens after an inquiry is sent?
The message reaches the editorial inbox and a person at this desk writes back by email. Turnaround depends on real workload; no auto-responder pretends otherwise.
Can the essays be cited or republished?
Short quotations with a link back are welcome. For reprinting a full essay elsewhere, write first — the terms page describes the details.