←BACK TO BLOG
>

Regime Change Is a Clustering Problem, Not a Feeling

QuantTradingVolatility
JUL 2026·6 MIN READ

#THE CHEAPEST SENTENCE IN TRADING

Every trader I know has said it after a bad month: the market's in a new regime. It's the most convenient sentence in the business: it explains a drawdown without admitting a mistake, and it can't be checked, because nobody defined what a regime was before the P&L went red. I've reached for it myself, and the thing that stopped me wasn't discipline, it was noticing the sentence never came with a number attached.

That's the tell. A claim that only ever surfaces after the fact, with no variable you could have measured beforehand, isn't a market observation. It's a description of your own trade going wrong, narrated in the third person.

#REGIMES THAT ARE MODELS, NOT MOODS

Two Sigma's 2021 paper, "A Machine Learning Approach to Regime Modeling," runs a Gaussian Mixture Model over the factors in their Factor Lens and comes out with four clusters: Crisis, Steady State, Inflation, and Walking on Ice. Nobody decided in advance that WOI should exist; the algorithm found a cluster of factor means, volatilities, and correlations that didn't look like the other three, and got a name after the fact. That output then feeds Venn's extreme-scenario tool directly, which means the classification isn't a narrative, it's an input other risk decisions get built on.

This isn't new territory. Hamilton's 1989 Econometrica paper treated regimes as a first-order Markov chain with transition probabilities estimated from data, not a story told once growth turned negative. More recent work pushes the same discipline across different math: a 2021 arXiv paper clusters monthly realized covariance matrices with agglomerative hierarchical clustering and checks the result against a VLSTAR switching model, while another applies Wasserstein k-means to cluster return distributions by shape rather than label. Different tools, same rule: define the state space before you know which state you're in.

#THE VARIABLES, NOT THE VIBE

If "regime change" is going to mean something, you need to specify in advance what you'd measure: the same list every model above actually uses, not whatever felt different this week.

  • ▸Factor means across the exposures you actually trade: equity beta, rates, credit, commodity, currency
  • ▸Realized and implied volatility, including the rate of change, not just the level
  • ▸Pairwise and average correlations across your book, since correlation breakdown is usually the real mechanism
  • ▸Cluster distance or transition probability from a model fit on history, not eyeballed after the drawdown
  • ▸A threshold for "different enough," written down before the month started

Volmageddon in One Chart

29.06Feb 1Feb 6Feb 9

The VIX closed at 37.32 on February 5, 2018 (up 115% in a single session from 17.31, its largest one-day spike on record) while the S&P 500 fell 4.1% that day, the mechanical signature of a real regime break against an otherwise calm week.

#WHEN IT'S REAL: FEBRUARY 5, 2018

Volmageddon is the cleanest regime change on record because it comes with numbers attached. The VIX closed at 17.31 on February 2 and 37.32 on February 5 (up roughly 115%, the largest one-day percentage move since the index launched in 1993) while the S&P 500 fell 4.1%, its worst day since 2011. The XIV ETN went from about $1.9 billion in assets to about $63 million in a single session and was delisted days later. That's not a vibe shift; it's a mechanical, instrument-level blowup with a specific transmission mechanism (short-vol ETN rebalancing feeding back into futures) that a clustering model over volatility and correlation would have flagged as an outlier state, not a normal Tuesday.

#WHEN IT ISN'T: AUGUST 2007

The Quant Quake is the harder case, because the broader market barely moved (equities and fixed income were calm) while statistical arbitrage funds got hit hard and in tandem: Goldman's Global Equity Opportunities Fund lost more than 30% in a week, and Tykhe Capital's largest fund lost about 20%. It would have been easy to call this a regime change in equities and stop there. Khandani and Lo's NBER paper did the opposite: they went into factor and transaction data and showed it was a deleveraging and correlation shock specific to crowded quant strategies, drawing an explicit parallel to 1998 LTCM, a measurable mechanism, even though the surface-level market data gave almost no reason to suspect one.

#THE TEST, NOT THE STORY

If regime change is going to license trading differently (cutting risk, flipping a signal, abandoning a strategy), it needs to clear the bar that Two Sigma's GMM, Hamilton's Markov chain, and Khandani and Lo's factor analysis all clear: a specification set before the fact, applied to data, producing a number you can be wrong about. Otherwise you're not describing the market, you're describing your P&L and giving it a more respectable name.

The honest version of "I think we're in a new regime" is "here's the cluster distance, here's the transition probability, here's the correlation matrix that moved." If you can't produce those, what you actually have is a bad month looking for an alibi.

Four Regimes, Not a Feeling

illustrative
Crisis
Steady State

Stylized illustration of an unsupervised model (like Two Sigma's Gaussian Mixture Model over its Factor Lens) separating market states by measurable variables: Crisis and Steady State sit at opposite corners of factor volatility and cross-factor correlation, while Inflation and Walking on Ice would occupy the space between.

#Further Reading