Regime attribution: making robustness inspectable

"Robust" might be the most abused word in systematic trading. Every deck claims it; almost none defines it. An aggregate track record — however long, however smooth — answers only one question: how did this do, on average, over the specific sequence of markets it happened to live through? The question an allocator actually needs answered is different: where does this break?

Regime attribution is how we make that question answerable — for ourselves first.

The decomposition

Every trading day in our history is tagged along several largely independent dimensions of market state:

Volatility state — realized and implied levels, and, importantly, the transitions: expansion days behave nothing like high-but-stable days.
Trend state — trending versus mean-reverting conditions in the traded underlyings, at multiple horizons.
Correlation state — whether cross-asset and intra-market correlations are compressed, normal, or in the kind of breakdown that turns hedges into positions.
Liquidity state — spreads, depth, and turnover in the instruments we actually trade, measured from our own captured data rather than an index proxy.

Strategy performance is then decomposed across these tags: return, downside ledger, exposure, and execution quality, per state. Not a factor regression with overlapping loadings — a direct, inspectable partition of the record: these are the high-vol trending days, and this is exactly how the strategy behaved on them.

What it exposes

Three uncomfortable discoveries this framework forces, sooner or later:

Concentration of edge. Some strategies earn their entire record in one corner of the state space — say, calm markets with compressed correlations — and pay rent everywhere else. That is not necessarily disqualifying, but it changes the strategy's role: it is a regime bet wearing a diversifier's clothes, and it must be sized and hedged as what it is.

Conditional fragility. A drawdown statistic computed over the full record can hide the fact that all of the pain arrives in one state — and that the state tends to persist. A 10% maximum drawdown means something very different when its entire mass sits inside correlation-breakdown weeks.

Missing evidence. Sometimes the honest answer is that a strategy has simply never traded through a given regime at meaningful size. Attribution makes that visible instead of letting the aggregate imply an experience the record does not contain. Absence of a stress test is not robustness; it is an untested assumption with a start date.

From attribution to action

Attribution is not a reporting exercise. It feeds decisions directly:

Sizing is set against a strategy's worst state, not its average day.
Portfolio construction pairs strategies whose weak states do not coincide — which requires knowing the weak states, not inferring them from a correlation of monthly returns.
Monitoring watches the live state tags. When the market walks into a region where a strategy has thin or poor evidence, exposure comes down before performance has to prove the point again.

The diligence angle

When we say robustness is inspectable, this is what we mean concretely: the record can be cut along market states, and the cuts are part of what we show — including the corners where the evidence is thin or the performance is poor. A manager who will only show you the aggregate is asking you to average away exactly the information you are there to find.

The market will run the decomposition eventually. The only choice is whether it happens on a research schedule or an allocator's statement.