A signal is usually defined in fixed terms: a 3% drop, a moving-average crossover, a close above a 20-bar high. The market around it is not fixed. The same event can be rare and informative in one environment and routine noise in another. This article shows why, with one signal compared across four market regimes, and what to change in your research as a result.
The same number, different events
Take a simple signal: "a single bar falls 3% or more." On its face it means the same thing every time. But how unusual a 3% fall is depends on how much the market typically moves.
If a market's typical daily move (one standard deviation) is 0.8%, a 3% drop is nearly four standard deviations. It is rare, and something has probably changed. If the typical daily move is 3.5%, the same drop is less than one standard deviation. It happens most weeks and may carry almost no information.
The table below makes this concrete with illustrative volatility levels for four regimes.
| Regime (illustrative) | Typical daily move (1 s.d.) | A 3% drop in s.d. units | How often a move this size occurs |
|---|---|---|---|
| Steady trend | 0.8% | 3.75 | Rarely |
| Range-bound | 1.0% | 3.0 | Occasionally |
| Drawdown & recovery | 2.0% | 1.5 | Regularly |
| Volatility shock | 3.5% | 0.86 | Most weeks |
Same rule, same threshold, four different events. A strategy that treats them as one is mixing together very different situations.
Volatility clusters, so context persists
If calm and turbulent days were scattered at random, context would matter less: each bar would be a fresh draw. They are not. Rama Cont's widely cited survey of the statistical properties of asset returns (2001) set out a list of "stylized facts" seen across many markets, including nonlinear dependence of returns over time. The best known of these is volatility clustering: large moves tend to be followed by large moves, and quiet periods by quiet periods. A 2023 study that retested Cont's list on intraday data for Dow 30 stocks found conclusive evidence for eight of the eleven facts, and no support for the remaining three.
For signal research, clustering has a practical consequence. The regime you are in today is informative about tomorrow, so a signal's meaning on a given bar depends on the bars before it. A test that averages across regimes hides that dependence.
One signal across four regimes: an illustrative comparison
To see the effect, imagine running the "3% single-bar drop, buy and hold for 5 bars" rule on each of four synthetic scenarios with identical parameters, fees and position size. The chart compares the average 5-bar return after the signal with the typical 5-bar move in each scenario.
- Average 5-bar return after signal
- Typical 5-bar move (1 s.d.)
Illustrative numbers, not market data. Typical move is one standard deviation of 5-bar returns in each scenario.
Show the numbers
| Average 5-bar return after signal | Typical 5-bar move (1 s.d.) | |
|---|---|---|
| Steady trend | 1.2% | 1.9% |
| Range-bound | 0.8% | 1.6% |
| Drawdown & recovery | -0.3% | 3.8% |
| Volatility shock | -1.4% | 6.2% |
Read the two series together, not separately.
- Steady trend: the signal fires rarely, and the average rebound of 1.2% is large compared with a typical 5-bar move of 1.9%. Here a sharp drop is a temporary break in an orderly trend.
- Range-bound: a smaller rebound of 0.8%, still meaningful against a typical move of 1.6%. The range pulls price back toward the middle.
- Drawdown & recovery: the average is slightly negative and tiny compared with a 3.8% typical move. The signal is close to noise.
- Volatility shock: the average is −1.4%, but the typical move is 6.2%. The signal fires often, the outcome is mostly variance, and buying the drop tends to catch further falls.
If you pooled all four scenarios, you might get a small positive average and conclude the rule "works a bit". In fact it works in two regimes, fails in one and means nothing in another. The pooled number describes no market that actually exists.
Three ways to put context into a hypothesis
Knowing that context matters is only useful if it changes how you write and test ideas.
1. Scale thresholds by volatility
Instead of "a 3% drop", define the signal as "a drop larger than 2.5 times the 20-bar standard deviation of returns." The threshold then adjusts automatically: it requires a bigger move in turbulent markets and a smaller one in calm markets, so the signal keeps roughly the same rarity. You can then check whether its meaning stays steadier across regimes.
2. State the regime in the hypothesis
If you believe a rule works only in ranges, say so: "In a range-bound market, after a drop of more than 2.5 s.d., the 5-bar return is positive on average, net of costs." Then test it in the other scenarios too, expecting it to fail there. A rule that fails where you predicted is more convincing than one that has only been tested where it was meant to work.
3. Size for context, not only for signal
Position size affects how a regime feels as much as entries do. Research by Moreira and Muir on volatility-managed portfolios found that, for several equity factors and the currency carry trade, changes in volatility were not offset by proportional changes in expected returns, so reducing exposure when volatility was high improved risk-adjusted results in their sample. That is a finding about specific portfolios and periods, not a rule. But the research question carries over: does your result change if position size shrinks as volatility rises? It is worth one extra run in your notes.
A checklist before trusting a signal result
Before you accept any test result, ask:
- What was typical volatility in the test period? Express the signal threshold in standard deviations, not just percentages.
- How many signals came from each regime? A result driven by one turbulent stretch is a result about that stretch.
- Does the sign of the result change between regimes? If it does, the pooled average is misleading.
- Do costs matter more in some regimes? A rule that trades often in volatile periods pays fees and slippage more often.
- Is the regime part of the hypothesis? If not, write it in and test again.
Context is part of the signal
A signal is not just the event. It is the event measured against the market it happened in. Once you start writing thresholds in volatility terms and naming the regime in each hypothesis, many "inconsistent" results become understandable. The rule was not unreliable; it was being asked to mean the same thing in different places.
To write those regime-aware hypotheses with clear failure criteria, see from trading idea to testable hypothesis. And when a summary of a cross-regime test tells you why a rule worked, what an AI trading bot can explain—and what it cannot helps you check whether that explanation is supported.