Can a smart model time the market's regimes? We killed the idea in an afternoon.
Grade B — pre-registered kill criterion, tested on both eras, decided before any model was built.
Status — FAILED (killed before modelling) · score v1
The claim
Most strategies get wrecked in bear markets. So the seductive idea is to add a regime model — a statistical brain (a Hidden Markov Model is the fashionable choice) that reads the market's mood from returns, volatility, breadth, and the VIX, keeps you out during the crashes, and — the real prize — gets you back in near the bottom, earlier than a plain moving-average rule would. Beat the simple "only trade when the market is above its 200-day average" gate, and you'd have something valuable.
The verdict, in one line
There is no prize to win. Before building any model, we computed the absolute ceiling — the most an ideal regime-timer could possibly add — and in the out-of-sample era that ceiling was negative. No model, however clever, can extract an edge that isn't there. We killed it in about fifteen minutes of compute, exactly as planned.
How we tested it (the cheap-kill discipline)
This is the part we're proudest of. Instead of spending weeks building and tuning an HMM, we pre-registered a kill criterion and ran a "ceiling computation" first:
- Take every trade our strategy would have made but the 200-day gate blocked (because the market was below its average).
- For each, ask: how many days before the market's next recovery did it fire?
- Then grant a perfect, impossible, look-ahead regime-timer that re-admits every one of those blocked trades at the ideal moment, and add up the profit it would have captured.
That sum is the ceiling — the best any regime model could conceivably do. Written in advance: if the ceiling doesn't clear a small bar, no model gets built.
What we found
Out-of-sample (2020–2026), the ceiling was negative at every horizon — even with perfect, cheating foresight:
| How early the perfect timer re-enters | Ceiling uplift | vs baseline |
|---|---|---|
| ≤ 30 days before recovery | −0.052R | +0.832R |
| ≤ 90 days | −0.064R | +0.832R |
| ≤ 180 days | −0.083R | +0.832R |
In the training era it was barely positive (+0.03 to +0.04R) but under the pre-registered bar — and remember, that's with perfect foresight. A real model captures only a fraction of a ceiling. When the ceiling itself is zero-to-negative, the game is over before it starts.
Why — and this is the real insight
The per-episode breakdown tells the whole story. The 2020 COVID rebound was a massive missed opportunity — the gate kept us out through a huge recovery. If that were the only kind of episode, a regime-timer would be gold. But the 2022 bear-market rallies were exactly the opposite: sharp, convincing bounces that the gate correctly kept us out of, saving large losses.
Here's the trap: the violent recovery and the vicious bear-rally look identical from the inside. Both are "high-volatility surges while the market is below its average." An earlier-re-entry model that catches the 2020 rebound necessarily also walks into the 2022 bear traps — they're the same statistical state, and with only a handful of episodes in history there's no reliable way to tell them apart in real time. The gate's cost (late re-entry after real bottoms) is almost exactly cancelled by its benefit (dodging fake ones). There is no free pool of profit for any earlier-re-entry model to capture — not an HMM, not a breadth signal, not anything.
How the score breaks down
| Dimension | Score | Why |
|---|---|---|
| Edge | 2 / 25 | The best-case ceiling is negative out-of-sample. There is no edge to capture, by construction. |
| Robustness | 5 / 25 | The negative result is robust: it holds at every re-entry horizon and even under perfect foresight. |
| Practicality | 3 / 25 | Nothing to implement — the idea was killed before a model existed. The simple 200-day gate already does the achievable job. |
| Evidence | 20 / 25 | Pre-registered kill criterion that fired in both eras exactly as designed; reproducible; a decisive upper-bound argument rather than a fragile backtest. |
| Total | 30 / 100 | No demonstrated edge — and, unusually, we can prove the ceiling is empty, not just that one model failed. |
Evidence Grade B: high confidence, because we bounded the entire idea rather than testing a single implementation. The one caveat is that the bound rests on a short history of regime episodes.
What it means for you
Two takeaways. First, "add an AI regime model" is one of the most common upgrades sold to systematic traders, and here the entire category is provably empty for this purpose — the profit you'd hope to capture doesn't exist to be captured. Second, and more useful: you can often kill an expensive research idea cheaply. Computing the best-case ceiling before building anything saved us weeks of work on a model that was mathematically doomed. That discipline is worth more than any single strategy.
References
- Hidden Markov Models for market-regime detection are a standard quantitative-finance topic; the novel step here is the pre-build ceiling computation as a falsification tool.
- The four-dimension scoring rubric behind the number above: The Tapelab Score.
Scored under score v1, the fixed public rubric summarised on the research index. Report last updated 2026-07-03; published from the research repository's canonical write-up, unedited. Nothing here is investment advice.