A regime is a period during which the market behaves in a characteristic way: trending or ranging, volatile or calm, rising or falling, with interest rates rising or falling. Regimes last months or years and change without warning. A strategy that depends on one regime will have a record that looks like skill while the regime lasts and like failure when it ends.
The common dependencies
Strategies that buy dips or hold long positions in stock indices depend on a rising market; they lose in a bear. Trend followers depend on trends existing; they bleed in ranges. Mean-reversion strategies depend on ranges; they are run over in trends. Carry strategies in currencies depend on calm; they lose in risk-off episodes. Volatility-selling strategies depend on nothing happening; they lose everything when it does.
Why records mislead
Most published records cover a few years. The 2010s were a long rising market with low rates and low volatility, and a strategy that was long equities did well for a decade without an edge. A record from 2020 to 2021 covers a crash and a recovery. A record from 2022 covers a rate shock. Each period rewards a different dependency, and a strategy's record from any one of them says little about the others.
Testing across regimes
The honest test is to look at the strategy's results year by year, and across periods known to differ: a bear year, a rate-rise year, a low-volatility year, a crisis. A strategy that was positive in most of them, and whose losing years are explained by a stated dependency rather than a surprise, has been tested. A strategy whose entire edge came from one period, and which was flat or negative outside it, has a regime, not an edge. Separating results before and after the strategy was developed is the sharpest version: a strategy that only works on recent data was fitted to recent data.
A strategy shows 21 per cent a year over 2023 to 2025. Its results from 2015 to 2022, once run, show 2 per cent a year with an 18 per cent drawdown in 2022. The strategy is long equities with a filter, and the filter added little. Its record is the record of the market it was run in. A second strategy shows 12 per cent a year over the same recent period and 9 per cent over the earlier one, with 2022 flat. The second has something the market did not give it.
Ask what regime an EA's record was produced in, and what it would do in the others. A vendor who can say "it is long-biased and will lose in a bear market, here is 2022" is describing a real strategy with a known dependency. A vendor who says it works in all conditions is describing a strategy that has not met the conditions it does not work in. The longer and more varied the test period, the less the record depends on one kind of weather.