Profitable in One Regime Means It's a Bet on the Regime, Not a Strategy
Published 2026-07-30 · By Drew Shelem
Why a confirmed edge can turn out to be a disguised bet on volatility — and why ordinary validation doesn’t catch it
A new article in the cycle. The last one explained what a real edge looks like: small, quiet, slow to confirm. But there’s a catch that slips through all of validation and shows up only in live trading, when it’s already too late. Your edge can be real — and still be not a strategy but a disguised bet that the market stays in the same state it was in during the backtest. This article is about telling a strategy from a bet on the regime, and why the cycle’s usual gates are powerless here.
Where this topic even comes from
A familiar plot, but with a new ending. You honestly validated a strategy: you gathered a sample, ran the out-of-sample fan, checked Monte Carlo, accounted for inertia. Everything passed, the edge is significant. You launch the bot — and for a month it works, then the market calms down, and the bot starts quietly losing. Nothing broke: the code is the same, the signal is the same. What changed is the market — or rather, its state. And your edge, it turns out, lived only in one state.
Here’s the crux, and it’s unpleasant. The edge you measured is an average over the market states that happened during the backtest. If those months were mostly stormy, your “edge” is a bet that the storminess continues. A calm comes — and the edge vanishes or reverses, though you changed nothing. And the most insidious part: ordinary validation doesn’t see this, because it tests transfer to new data, not to a different regime. Let’s take apart how this works and how to catch the catch before it catches you.
Step 1. What a regime is and why strategies depend on it
A regime is a stable state of the market that holds for a while and then switches. The most important regime on these markets is volatility: there are stormy periods, when the price moves widely and sharply, and calm ones, when it’s sluggish and marks time. We already met this property in the article on how losses come in runs: volatility clusters — stormy sticks to stormy, quiet to quiet.
Why strategies depend on the regime — with an example. Take a strategy that catches breakouts: it enters when the price has moved sharply, betting the move will continue. In a stormy regime this works: sharp moves have momentum, the breakout continues, you’re in the black. In a calm regime the same strategy is a loss: a sluggish market dampens moves, the breakout turns out false, the price pulls back, and you get cut. One and the same logic brings profit in one regime and loss in another. So its result is not a property of the strategy but a property of the regime it traded in.
Step 2. How a bet on the regime disguises itself as an edge
Now watch how this fools validation. Take that same strategy: in the stormy regime it brings +3 cents per trade, in the calm one it loses 2 cents. And run it over a period where the market was stormy 70% of the time — an ordinary situation if you got a volatile slice of history.
What will the backtest show? It averages over the whole period: 70% of trades at +3 cents and 30% at −2 cents give a blended edge of about +1.3 cents per trade. Let’s run the statistics:
| What we look at | Value |
|---|---|
| Blended edge | +1.34 cents/trade |
| Confidence measure (t) | +4.81 |
| Naive validation verdict | significant, the strategy “is profitable” |
Stop here. A confidence measure of 4.81 is a confident “the edge isn’t random,” twice the threshold. The strategy passes the significance check with room to spare, looks real and profitable. You’d launch it. And you’d launch a bet on the regime, mistaking it for a strategy.
Step 3. The regime split reveals the catch
How to catch the deception? Split those same trades by regime and look at the edge in each separately. Here’s the same backtest, but split:
| Regime | Edge per trade | Confidence measure | Trades |
|---|---|---|---|
| High volatility | +2.76 cents | +8.39 | 2118 |
| Calm volatility | −2.07 cents | −4.06 | 882 |
The picture changes instantly. The whole plus sits in high volatility; in the calm the strategy doesn’t just fail to earn, it confidently loses (a confidence measure of −4.06 is a significant minus). The blended “edge” of +1.3 was an illusion of averaging: it exists only as long as the sample has enough stormy trades to cover the loss from the calm ones.
This is the very definition of a bet on the regime: the edge lives in one regime and dies (or reverses) in another. Your strategy doesn’t predict the outcome — it bets that the market stays stormy. While it’s stormy, you’re in the black; when it turns quiet, you’re in the red, and no bot-tuning will change that, because the problem isn’t in the bot but in the fact that the regime switched.
Step 4. A regime shift kills the edge — and ordinary validation doesn’t see it
Let’s check on a live scenario. Suppose after launch the market calmed down, and now high volatility is only 30% of the time (it was 70%). The same bot, the same code — but the regime mix flipped. Let’s average over many such periods, so as not to depend on one lucky one:
| Live: high volatility 30% of the time | Value |
|---|---|
| Average edge | −0.53 cents/trade |
| Share of periods in the red | 97% |
A “validated,” significant, backtest-profitable strategy loses in live trading almost always. Nothing broke — the regime simply switched, and with the regime mix the sign of the edge flipped too.
And here’s the most important thing, the reason this article exists. The cycle’s usual gates don’t catch this catch. If both the training slice and all the out-of-sample slices are taken from one stormy period — the out-of-sample fan will show a plus, Monte Carlo will show “robust,” because they test transfer to new data of the same regime, not to a different regime. You can pass all of validation perfectly and still launch a bet on the regime. The only thing that catches it is an explicit split by regime: seeing whether the edge survives separately in a stormy and separately in a calm market.
What would refute this — and how to test your strategy
The condition on which the thesis would collapse (and didn’t). If the regime split did not change the picture — if the edge were the same in a stormy and a calm market — then there’d be no regime dependence, and the blended edge would be an honest strategy. But the split reveals +2.76 against −2.07: the edge is entirely in one regime. And a regime shift zeros out the result 97% of the time. Both attempts confirmed the thesis.
How to test your strategy — three steps.
- Define the regime and tag each of your trades: did it happen in a stormy market or a calm one (the simplest measure of storminess is the recent spot volatility at the moment of entry).
- Compute the edge separately in each regime. If it exists only in one, and in the other it’s negative or zero — you have a bet on the regime, not a strategy.
- Decide consciously. A bet on the regime isn’t necessarily bad: if you know you’re trading only in a stormy market and turn the bot on only in that regime, it’s a legitimate conditional strategy. What’s bad is when you don’t know it and launch it blind, mistaking it for all-weather.
What I don’t know, and where the limits are
- A regime isn’t only volatility. There are regimes of trend versus range, time of day, liquidity, the news backdrop. I took volatility as the most important and vivid one, but you should split by every regime you can define.
- Defining the regime is itself a modeling choice. Where the line between “stormy” and “calm” falls — you decide, and the split depends on that line. It’s another degree of freedom, and it can’t be tuned to the result (or you’re back at the overfitting from the article on finding an edge).
- You can’t test a regime that wasn’t in the sample. If your whole backtest is one regime, you fundamentally don’t know how the strategy will behave in another until it arrives. This can’t be fixed by computation — only by caution and a small bet size until you’ve seen the second regime.
- The numbers are illustrative, from synthetics. The per-regime edge sizes and mixes were set by me; on your market they’re your own.
What to do tomorrow
- Always split your trades by regime before believing the blended edge. One overall plus may be the average of “I earn in the storm” and “I lose in the calm.” Until you’ve split it, you don’t know what you have.
- Require the edge to survive across regimes — or know it’s a conditional bet. An all-weather strategy earns in both; a regime one only in one, and it should be turned on only in its regime.
- Don’t trust the out-of-sample fan and Monte Carlo on the regime question. They catch transfer to new data, but not to a different regime; regime dependence is caught only by an explicit split.
- Until you’ve seen the second regime — keep the bet small. A strategy validated in one regime is reasonably run at a small size until the market shows you another regime and you’ve checked it there.
- Accumulate trades across different regimes before delivering a verdict. You can run a strategy long enough to catch both storm and calm, without risking money, in paper mode — free, before any subscription.
FAQ
The strategy passed all of validation — how can it be a bet on the regime? Easily: if the whole backtest fell in one regime, all of validation tested transfer within that regime. The out-of-sample fan and Monte Carlo don’t ask “and what about another regime” — they never saw that regime. Only an explicit split by regime catches it.
How do I define the regime in practice? The simplest measure is the recent spot volatility at the moment of entry: above some threshold — a stormy regime, below — a calm one. Don’t tune the threshold to the result, take a reasonable one in advance. You can also split along other axes (trend/range, time of day).
Is a bet on the regime always bad? No. What’s bad is launching it blind, mistaking it for all-weather. If you know you earn only in a stormy market and turn the bot on only then, it’s an honest conditional strategy. The problem isn’t the dependence itself, but not knowing about it.
My edge is only in high volatility — what do I do? Realize it’s a conditional strategy, and trade it only in its regime, determining the regime at entry. And keep the bet modest, because regimes change suddenly, and at the switch you may find yourself in the wrong one.
Can’t I just add a volatility filter? You can — but remember that a filter cuts the number of trades (there was an article about finding an edge: a filter reduces the denominator), and that the filter’s threshold is a degree of freedom that can’t be tuned to the result. A filter is honest if its threshold is set by a reason, not by fitting.
Disclaimer
This material is educational and is not financial advice. Past results do not predict future results. Trading on prediction markets carries real risk, up to and including the total loss of capital. Access to Polymarket is restricted or prohibited in some jurisdictions — verify legality where you live before trading.