Why we run three models instead of one

The problem with one model

Any single regime model is a lens with a defect. A moving-average filter is transparent and robust, but it only sees price trend — it will call a slow-motion top "bull" all the way down the first leg. A statistical model like a Hidden Markov Model sees more — volatility, cross-market stress — but it's a black box to most users, and black boxes fail silently.

The uncomfortable truth of quantitative finance is that model risk doesn't diversify away inside one model, no matter how sophisticated. It diversifies across genuinely different models.

Three lenses, deliberately unlike each other

Regime Compass runs three classifiers per market, chosen for how differently they fail:

• A 3-state Hidden Markov Model — probabilistic, regime-switching, driven by the joint behaviour of returns and volatility. Fast to smell stress, occasionally jumpy.
• Simple moving-average filters — the sixty-year-old workhorse of trend following. Slow, transparent, nearly impossible to fool for long.
• Exponential moving-average filters — the same logic with more weight on recent prices. Faster to turn, at the cost of more whipsaws.

None of these is secret — the value isn't in any one model's cleverness. It's in the structure of the disagreement.

Agreement is conviction; disagreement is information

When all three call the same regime, the classification is about as reliable as this kind of inference gets — that's when our blended risk score reads decisively. When they split, that's not a failure of the system. It's usually the most interesting moment on the board: turning points look exactly like a fast model flipping while a slow one holds.

A unanimous board tells you what the market is. A split board tells you what it might be becoming.

This is why the site shows the models side by side instead of averaging them into false precision. The blended gauge summarises; the disagreement underneath is where a careful reader looks next.

The check we impose on ourselves

Multiple models create a temptation: quietly lean on whichever looked best recently. We avoid it by auditing the primary model walk-forward — refit quarterly on data available at the time, signals applied the next day, costs included — and publishing the results, including the markets where it loses to buy-and-hold. A model you're allowed to see fail is worth more than one that's never wrong on its own website.