Is my volatility regime just telling me the time?

Intraday volatility has a shape you could set a watch by: quiet through the Asian session, busy at the London and New York overlaps, quiet again. It is one of the most reliably documented patterns in market data.

Anything you compute from a volatility path inherits that shape. A market-state label built on intraday data is partly a reading of the clock, and the share that is clock rather than market is measurable. I measured it on six of my own features and the answer was larger than I expected.

I call a label in that condition a Clock Regime.

The measurement

For each feature, the per-bar label is derived twice over 98,864 hourly bars of one FX pair: once from the series as it stands, and once after removing the hour-of-day pattern from the feature’s own input — dividing each observation by the statistic of the same hour over prior completed trading days only, so nothing reads its own future.

Then: what fraction of labels changed?

All six moved. Not one of them was measuring only the market.

The result that stopped me

The volatility regime classifier labels each bar CALM, COMPRESSION, EXPANSION or UNSTABLE. Here is what removing the clock did to it:

COMPRESSION went to zero. Not smaller — gone. All 8,629 bars it had flagged became calm, along with most of the unstable and expansion bars.

On this pair, over this history, the classifier’s non-calm states were substantially the intraday cycle. It was, for the most part, telling me the time.

Why the obvious test would have missed this

The natural way to check a feature for time-of-day dependence is to correlate its output against the hour. I ran that first, and on three of these six features it returned one to two per cent — negligible, dismissable, move on.

Those same three features moved on ten to forty-three per cent of observations.

The two measurements answer different questions. Correlation asks could you guess this label from the clock? Churn asks would this label have been different in a world without one? A feature that smooths or aggregates can score near zero on the first while carrying the effect at full strength — averaging over fourteen hours flattens the hourly signature without removing what it did to the level.

If you have only ever run the correlation, you have not tested this.

And a daily feature is not immune

Three of the six emit a label that never changes within a day. Their correlation with hour-of-day is zero by construction — there is no intraday variation for the clock to explain. That looks like immunity.

Corrected at the input, one moved on 14.9% of days. Another kept the same number of flagged events while roughly a third of them became different days.

The mechanism is the aggregation itself. Squared returns are as free of the clock as returns; summing them over a day is what puts the hour back in. Any statistic that sums or averages within-day quantities inherits the intraday profile no matter how clock-free its inputs are.

What I did to keep the number honest

Every family carries a control that isolates the correction from the mechanics of applying it.

Removing the pattern needs a history to build the profile from, so the earliest observations have none and are dropped — and dropping observations can shift a rolling window or re-phase a block all by itself. So each family is compared against a placebo: the uncorrected input, re-derived over the corrected arm’s exact observation set. Whatever that moves is the mechanics, not the correction. On the volatility forecast it moved 0.05%. On the jump flag, zero.

For the two features that consume whole bars rather than returns, the correction requires rebuilding the bar series, so those are compared against a raw rebuild put through the identical reconstruction. And the rebuild is validated rather than assumed: the raw reconstruction reproduces the label distributions measured independently on the original bars, to three decimals — 0.8811 against 0.881, and 0.6143 against 0.614. A rebuild that had distorted the series would not land on those.

What to do with this

Ask what your volatility state is made of. If it is computed from an intraday path and nobody removed the daily shape, some of it is the clock, and the fraction is not small enough to wave away.

Do not accept a correlation against the hour as the test. It is the wrong measurement and it will clear a feature that moves by forty per cent.

Deseasonalise before you condition, not after. A state used to slice returns into “compression” and “calm” buckets carries the clock into every bucket it makes — and a comparison between those buckets is then partly a comparison between times of day.

Limits

One pair, one grain, one history. A single-instrument measurement establishes nothing about any other, and the size of the effect will depend on the pair’s own session structure.

Nothing here says these features are worthless — a corrected feature is still a feature, and the correction is a few lines of arithmetic over a profile you already have the data to build. It says that the uncorrected version is measuring something other than what its name claims, by an amount that is worth knowing before you condition anything on it.