A strategy can work repeatedly and then stop working without its rules changing. Buying a dip may succeed while temporary selling is absorbed and fail when lower prices reflect a lasting change in value. The pattern on the chart can look similar; the mechanism underneath it differs.
“The regime changed” is a useful explanation only if it identifies something observable. Otherwise it can become a story attached to losses after the fact.
Trend, volatility, and liquidity
Trend describes the persistence of direction. Volatility describes the scale of fluctuations. Liquidity describes the cost and feasibility of trading a quantity. These dimensions can change independently.
A market can trend steadily with modest daily movement. It can swing violently around the same average. It can appear quiet until a large order reveals that very little liquidity is available. Calling all three “risk on” loses information that matters for the trade.
For a trend-following approach, the question is whether directional moves persist long enough to cover reversals and costs. For a mean-reversion approach, the question is whether deviations tend to be temporary. Both require a reason to expect the relevant behavior at the chosen horizon.
The horizon of the evidence
Moskowitz, Ooi and Pedersen document persistence over one- to twelve-month horizons in a historical sample of 58 liquid futures, with partial reversal over longer periods. That result does not imply that a five-minute breakout should work. The holding period and population are part of the finding.
Economic information operates on different clocks too. A central-bank decision can change prices immediately while its effect on production and employment unfolds over months. A market can rise during weak reported activity if investors are revising expectations about what happens next.
Positioning data can help explain who might need to respond, but its timing and categories matter. The CFTC’s COT methodology ordinarily describes Tuesday positions released Friday; categories are based on business purpose, not a complete explanation of each position. Treating that report as a live map of intentions would overstate its information.
A condition defined in advance
Suppose a strategy buys short-lived declines on the premise that selling pressure fades. Define what would support that premise: recoveries within a stated interval, stable spreads and no new information that changes the underlying valuation. Define what would weaken it: repeated failed recoveries, a widening trading range or persistent selling after relevant news.
The thresholds need to be written before seeing the results. “A trend exists whenever the next trade wins” is not a usable classification. Nor is a rule that retrospectively moves the beginning of a regime to its most favorable price.
You can also separate classification from action. An uncertain change in conditions might justify testing a smaller exposure rather than making a confident forecast of a new cycle. The size of the response should reflect the uncertainty of the evidence.
A condition filter has costs too
Imagine a base strategy earns $200 over a sample. A filter avoids trades that would have lost $80, but it also skips trades that would have gained $120. The filtered result is $160 before any extra costs: $200 + $80 − $120.
Showing only the avoided losses makes the filter look better than it is. Record both sides, use information available at the decision time and assess the rule on later data after its definition is fixed.
In the example, the filter correctly avoided some losses and still reduced the total return. A regime label has to improve the decisions made with information available at the time, including the decision to skip a trade that would have won.