There is a version of regime detection that makes bold claims: "We know what the market will do next." That version is marketing. The version we use is more modest and, because of that modesty, more useful. It starts with an observation that is hard to dispute: markets do not behave the same way all the time.

The volatility of 2020 is not the volatility of 2017. The correlation structure of a credit crisis is not the correlation structure of a bull run. These are not random fluctuations around a single stable mean — they are structurally distinct environments, each with its own statistical character. The question regime detection asks is not what happens next, but which environment are we likely in right now.

The Model Behind the Classification

Cacao uses a Hidden Markov Model (HMM) trained on a set of nine market indicators. An HMM is a probabilistic model that assumes the observed data — the indicators — are generated by an underlying sequence of hidden states that you cannot observe directly. The model learns the statistical fingerprint of each state and, given new data, outputs a probability distribution across states.

The model doesn't say "we are in a bull market." It says "given current evidence, the probability of being in each environment is: X%, Y%, Z%."

That distinction matters. A point prediction pretends to certainty that doesn't exist. A probability distribution is honest about what the data can and cannot tell us, and it allows the portfolio optimization to respond proportionally rather than all-or-nothing.

The Nine Indicators

The model reads across market structure, momentum, credit conditions, and volatility. Each indicator captures a different dimension of market character. Together, they provide a more complete read than any single signal could.

Price Momentum
Trend direction and strength in equity indices over multiple lookback windows.
Realized Volatility
Short-term volatility of returns — a direct measure of market uncertainty.
Implied Volatility (VIX)
Forward-looking fear gauge derived from options pricing on the S&P 500.
Credit Spreads
The gap between corporate and Treasury yields — a proxy for credit stress.
Yield Curve Shape
The slope of the Treasury curve, historically predictive of economic regimes.
Breadth
Participation: what fraction of stocks are above their moving averages.
Cross-Asset Correlation
How assets are moving relative to each other — high correlation often signals stress.
Rate of Change
Acceleration or deceleration of key indicators, not just their level.
Liquidity Conditions
Bid-ask spreads and market depth as a measure of friction in the system.

The nine indicators processed daily by the regime classifier.

What the Classification Drives

The regime output doesn't directly pick stocks or time the market. It feeds into the portfolio optimizer as a context signal. Specifically, it influences two things: the expected return assumptions used in optimization, and the covariance matrix that describes how assets are likely to move together.

In a high-stress environment, correlations between asset classes tend to spike — diversification offers less protection than the long-run average would suggest. A covariance matrix estimated only on recent data captures this; one estimated on a long historical window does not. Regime conditioning allows the model to select the appropriate estimation regime for the current environment rather than averaging across all of them.

9
Market indicators read daily. The classifier runs each trading day on the latest data, updating the probability distribution across market states. No human intervenes in the classification — the signal is purely systematic.

The Honest Limitations

Regime models have a real and documented weakness: they can lag at turning points. When a market transitions rapidly — a flash crash, a pandemic onset, a banking shock — the model is looking at recent data that still reflects the previous environment. There is an inherent latency.

This is not a defect to be fixed so much as a property to be understood. The model is calibrated to avoid overreacting to short-term noise, which means it accepts some lag at transitions in exchange for stability during trending environments. The framing from Corey Hoffstein is apt here: the goal is a slight statistical edge applied consistently, not perfect prescience. A model that avoids most of the worst outcomes while participating in most of the best outcomes will compound well over time, even if it looks wrong at any given moment.

A slight edge, applied consistently and without interference, outperforms a large edge applied with discretion.

— Paraphrased from Corey Hoffstein, Newfound Research

Why This Matters for Individual Investors

The practical implication is not that Cacao will get you out before every downturn. It won't, and any system that claims to is selling something. The implication is that the portfolio responds to evidence rather than to anxiety. When indicators deteriorate, exposure adjusts. When they recover, exposure adjusts back. This happens systematically, without the behavioral drag of an investor trying to time the same transitions based on how they feel about the news.

The regime model is one part of a larger architecture. But it is the part that makes the system context-aware — that distinguishes between a portfolio optimized for one static view of the world and one that reads the environment and adapts to it. Not perfectly. But consistently, and without the interference that ruins most investors' long-run outcomes.


The content in this piece is educational and describes the general methodology used in Cacao's systematic approach. It does not constitute investment advice. Strategies are subject to change. Past classification accuracy does not guarantee future results.