A strategy can look robust across years of data and still have learned one particular market’s habits. It may thrive in quiet trends, then lose its footing when volatility spikes or liquidity thins. The useful question is not whether it works in every regime. Few strategies do. It’s whether you know which conditions pay for the strategy, which conditions make it stumble, and whether the pattern holds up beyond the sample that revealed it.
Start by treating regime dependence as a property to measure, not a story to tell after seeing the equity curve. Define a small number of market conditions using information the strategy could actually have known at the time, then inspect the strategy’s behavior inside each one.
How can I tell if my strategy only works in one regime?
Break results down by a few economically meaningful conditions: realized volatility, trend strength, liquidity, or funding direction. Keep the definitions simple enough to explain. For example, classify each day as high or low volatility using a trailing 60-day measure and a threshold calculated from earlier data. A full-sample median is tempting, but it lets the future help label the past.
Compare net returns, drawdown, turnover, and trade count across the groups. A positive average in one bucket means little if it comes from three trades. And a strategy may earn similar returns across buckets while taking radically different risks to get them.
| Measure | What it can reveal | What to check |
|---|---|---|
| Net return per trade | Where the apparent edge is concentrated | Trade count and uncertainty in each group |
| Maximum drawdown | Conditions that make losses cluster | Whether one episode dominates the result |
| Turnover and costs | Whether a regime makes execution more expensive | Fees, spread, impact, and funding together |
Here’s a small trap: a trend strategy can show weaker returns in high volatility simply because those periods contain both sharp trends and violent reversals. “High volatility” isn’t a complete explanation. Slice until the result becomes useful, then stop before every bucket becomes a new opportunity to find a flattering statistic.
How many regimes should I test?
Fewer than your curiosity wants. If you divide by volatility, trend, liquidity, weekday, and asset, you’ll quickly end up with a tiny number of trades in dozens of cells. The table looks sophisticated; the evidence has evaporated.
Pick one or two drivers with a plausible link to the strategy’s mechanism. A funding arbitrage might depend on funding sign and liquidity. A breakout system might depend on trend persistence and volatility. State those choices before checking the results, and report the sample size alongside each number.
Don’t build a “regime detector” by searching for the labels that maximize historical Sharpe. That is optimization with a nicer name. If a threshold is part of the strategy, choose it using training data and freeze it for the next period.
Does walk-forward testing prove an edge works across regimes?
No. Walk-forward testing asks whether a research process can make decisions on later data without using it to fit those decisions. It doesn’t guarantee that the later periods contain the conditions you care about, or that the results are precise enough to distinguish skill from noise.
Read each test window as a dated observation. Which regimes appeared? How many trades occurred in each? Did the strategy’s parameters change sharply when the training window moved? A good aggregate score can conceal a strategy that repeatedly gets its gains from one unusual stretch.
If you’ve tried many features, thresholds, and variants, account for that search. The prettiest regime breakdown is especially easy to overfit because there are so many plausible ways to divide a market.
What should I do when the strategy fails in a regime?
First decide whether the failure contradicts the mechanism. A market-making idea losing money when spreads collapse may be unsurprising. A supposedly market-neutral spread strategy taking large directional losses during a routine volatility jump deserves a closer look at its hedge, marks, or position sizing.
Then test the explanation with a change in data or design that could disprove it. Include realistic costs for the conditions that hurt: spreads and impact often widen precisely when a strategy wants to trade most. For perpetual futures, funding can change sign. For options, the price of a hedge can move with volatility and liquidity. A regime analysis that reports gross returns while ignoring those effects can point in the wrong direction.
If the strategy is explicitly designed to stand down, define the rule using information available at decision time. Measure the missed gains as well as avoided losses. A filter that removes every painful historical period may also remove the few periods that pay for the strategy.
Can paper trading tell me whether the edge survives?
Paper trading can test whether signals, orders, and costs behave as expected in the current market. It can expose operational gaps and show whether live conditions resemble the assumptions in the backtest. But a few quiet weeks cannot validate performance in a volatility shock that hasn’t happened yet.
Track the conditions observed during paper trading and compare fills, fees, funding, and missed orders with the backtest’s assumptions. Keep the claim proportional to the evidence: “execution matched expectations in this period” is defensible; “the strategy is robust to regime change” usually takes much longer to establish.
← All posts
