Why volatility regimes matter more than direction

If you watch a chart long enough, you start noticing two things layered on top of each other. The first is direction: the market is going up, down, or sideways. The second is texture— the bars are calm and orderly, or jittery and gapping, or one bar dwarfs the previous twenty. Most retail trading systems spend all their effort predicting the first thing. We'd argue the second thing matters more.

Direction is what people argue about on Twitter. Volatility regime is what determines whether your strategy makes money or loses your shirt.

The same trade in two different worlds

Imagine a simple long-only momentum bot: buy when the 20-day moving average crosses above the 200-day, sell when it crosses back. It's the most-blogged-about strategy in existence; not because it's great, but because it's easy to explain.

In a calm uptrend — bars don't move much, drift is positive — that strategy looks like a hero. Buy near the cross, ride the slow grind, exit on the way back down. The drawdowns are tiny because the underlying volatility is tiny. You can size up. You can use leverage.

Now run the same strategy through 2020's March or 2008's October. The bars are 5%, 8%, 10% wide. Crossovers happen on noise, not signal. By the time the system enters, the move is half-reverted. Stops get blown through on a single bar. The strategy that printed money in calm times becomes a mechanism for converting capital into commission.

The strategy didn't change. The market's volatility regime did.

Regimes are the hidden state

We borrow the word regime from regime-switching models in financial econometrics, but the intuition predates the math. Markets aren't one thing — they have moods, and the mood persists. Calm low-vol stretches last weeks or months. Crisis high-vol bursts last days. Inside each regime, the same set of features (returns, dispersion, autocorrelation, gap behavior) carries different information.

These bands aren't hypothetical. Pull the daily returns of SPY over the last fifteen years and bucket them by trailing 20-day realized vol. You'll find the same shape every time: a long left tail of calm days, a fat hump of typical days, and a thin tail of crisis days. The crisis days carry most of the dispersion in the underlying return distribution. That's not noise — it's structural.

Why direction prediction is harder than it looks

Predicting whether tomorrow is up or down is, in any honest sample of equities, a coin flip with a tiny edge from drift. The drift is real (~6-8% annualized for the S&P 500 over a long horizon) but it's buried under daily noise of 1% standard deviation in calm times and 3-5% in crisis. To consistently beat random on direction, you need either a) edge that compounds your way out of the noise, or b) a small enough sample that survivorship dominates.

Most retail backtests fall into category (b). They show a beautiful equity curve from 2013 to 2019 — a regime that was historically anomalous in its calmness — and then explode in March 2020, August 2024, October 2025, etc. The backtest didn't lie. It just measured the wrong thing.

What regime detection asks instead

Rather than “is tomorrow up or down,” we ask: what kind of market is this? Calm uptrend? Choppy sideways? Crisis crash? Recovering bear?

That question is much easier. Volatility regime persists. Once you're in calm territory, the next few days are usually still calm — high autocorrelation. Once you're in crisis, you're probably still in crisis tomorrow. So a regime classifier doesn't need to predict; it needs to recognize. And recognition over a few days' worth of bars is statistically tractable in a way next-day direction prediction isn't.

Then your strategy can adapt. In a calm regime: full allocation, leverage allowed, longer holding periods. In crisis: cut exposure, halt new entries, raise stop tightness. Same strategy, different sizing — driven by what the market is currently doing, not what you hope it will do.

Regimes don't replace alpha — they amplify or mute it

We're not claiming a regime classifier turns a bad strategy into a good one. If your underlying signal is noise, regime detection just makes you noisy in a calm sort of way. What regimes do is scale — they tell you when to lean in and when to step back, conditional on signal you already have.

Concretely: a strategy with a real but small edge will underperform if it's sized constantly across all market states. The same strategy, sized 2x in calm and 0.25x in crisis, will outperform itself by a factor that depends almost entirely on how persistent regimes are. For US equities, the answer is “persistent enough that it's a free lunch ignoring it.”

What this looks like at HMM Trade

Our bot runs a Hidden Markov Model that classifies every new bar into one of seven volatility regimes — from DEEP_BEAR (high vol, recent crashes) to TOP_BULL (high vol, but on the upside). The classifier doesn't care about direction directly; it cares about the joint distribution of returns + volatility + a few engineered features. The downstream strategy then uses the regime label to size positions, set leverage, and choose between active vs defensive substrategies.

We picked seven regimes not because the number is sacred but because — for the universes we trade — BIC selects 7 when given a choice between 3 and 8. (More on that in a future post.) Other universes might want 5 or 9. The principle is the same: enough granularity to differentiate regimes that matter, not so many that the model overfits.

The takeaway

The next time you read a backtest, ask: what was the volatility regime during the test window? If it was a calm uptrend, the result tells you the strategy works in calm uptrends. That's useful! It's also a fraction of what you'd need to know before risking capital.

Direction is a question about the future. Regime is a question about the present. And the present is the only thing you can actually observe.