Does a technical edge survive outside a crash? We split the data into calm and volatile
A signal that only works when the market is falling apart is a different product from one that works in normal times. So we separated the two.
Every backtest number is an average over whatever market happened during the test. That is a problem, because a signal that only "works" when the market is falling apart is a completely different product from one that works in normal conditions — and a single blended average hides which one you have.
So we split our measurement window into two regimes: a high-volatility stretch (late February to late June 2026, a genuinely turbulent period for Indian equities) and the calmer periods around it. Then we asked how our strong-setup tag behaved in each. The two answers were not the same.
| Regime | Horizon | Excess vs universe | t-stat | Days |
|---|---|---|---|---|
| Volatile | 3 days | −0.62% | −2.7 | 71 |
| Calm | 3 days | −0.21% | −2.0 | 237 |
| Volatile | 10 days | −1.17% | −1.7 | 71 |
| Calm | 10 days | −0.10% | −0.4 (ns) | 230 |
The underperformance lives in the crash
Read down the table and the pattern is clear. In the volatile stretch, the tag lagged sharply — by 0.62% over three days and over a full percent across ten. In calm periods the same tag was only marginally behind at three days and statistically indistinguishable from the market at ten. In other words, the raw aggregate we reported elsewhere is dominated by the crash window; strip that out and most of the effect thins to noise.
Is that a broken signal? Not necessarily
A momentum-style screen taking damage in a high-volatility sell-off is expected behaviour, not evidence of a flaw. Trend-following approaches structurally struggle when trends snap; that is the cost of the style, well known and well documented. What the split tells you is narrower and more useful than "the tag is bad": it tells you the tag's raw underperformance is a regime story, and the honest way to describe any such number is to name the regime it came from.
How we measured this
Daily excess over the scanned universe, computed separately within each regime window, with Newey–West standard errors over the day series. Regime boundaries are a judgement call based on realised volatility; the calm sample is larger, the volatile sample smaller and therefore noisier at the longer horizon.
The general lesson
Before trusting any backtested edge — ours or anyone's — ask "in which regime?" A number that looks decisive over a full sample can be one turbulent quarter wearing a confident face. Splitting by regime is one of the cheapest, most revealing checks you can run, and it is the difference between "this works" and "this worked, mostly during the crash."
Key takeaways
- Split by regime, the strong-setup tag's underperformance concentrates in the volatile window (−0.62% at 3 days).
- In calm periods it is marginal at three days and effectively zero at ten.
- A trend screen struggling in a crash is expected, not a broken signal.
- A full-sample average can be dominated by one turbulent stretch.
- Always ask which regime a backtest number came from before trusting it.
See the scanner these numbers come from
StockLearn runs this read on ~2,000 NSE stocks every evening. Nifty 50 is free, no login.
Browse today's scan →This article reports measured technical results from StockLearn's own scan history over a specific, limited period. It is educational research, not investment advice or a recommendation to buy or sell any security. Past statistical tendencies do not predict future returns.