Conditioning a Signal on the Volatility Regime
A signal's average IC over a full history can hide the fact that it only works in calm markets, or only in stressed ones — splitting the backtest by volatility regime exposes that and tells you when to trust the signal at all.
Prerequisites: Framing a Research Question
An average is a single number standing in for what might be two very different worlds. A signal's overall IC of 0.02 could mean it's a steady 0.02 in every market, or it could mean it's +0.08 when volatility is low and −0.02 when volatility spikes, averaging out to something that looks mediocre but is actually two distinct, tradeable facts. Splitting the backtest by volatility regime is how you find out which story is true.
Why regime matters here specifically
Volatility regime changes the character of the cross-section in ways that plausibly interact with almost any signal. In calm markets, price moves are more likely driven by fundamentals, which favours signals built on fundamentals. In stressed markets, correlations rise, liquidity thins, and forced selling can dominate — flow-driven, not information-driven, price moves that many fundamentally-motivated signals are not built to predict and that can actively work against them. A momentum signal, similarly, tends to behave differently in trending calm markets than in choppy stressed ones.
The test is simple: classify each period into a small number of volatility buckets (say, terciles of realized market volatility, computed without look-ahead — using only data available up to that point), then measure IC separately within each bucket.
A signal that is only positive in one volatility regime isn't necessarily broken — it may just need to be sized down or turned off in the regime where it doesn't work, rather than blended at a constant weight across all regimes.
A worked example
A quality signal's full-sample IC is 0.015. Splitting by realized-volatility tercile: low-vol months average 0.035, medium-vol months average 0.012, high-vol months average −0.010. The full-sample number was a misleading blend of a genuinely useful signal and a period where it actively hurt. A researcher armed with this split can build a simple regime-conditional rule — scale the signal's weight down, or to zero, when trailing realized volatility crosses into the top tercile — and expect a better risk-adjusted outcome than trading the signal at constant weight through all regimes.
Use a regime classifier that only looks backward at the time of each decision. Defining "high vol" using the full-sample distribution and then testing on the same sample is a subtle look-ahead — the regime label itself becomes informed by data the strategy wouldn't have had live.
Related concepts
Practice in interviews
Further reading
- Ang, Asset Management: A Systematic Approach to Factor Investing (ch. on regimes)