30 / 97
Epps Effect
Working definition
The systematic decline of measured correlation between two price series as the sampling interval shrinks; at fine grains the instruments rarely print at the same instant, and the correlation formula reads that asynchrony as independence.
A correlation number is a claim about a specific sampling interval: measured at this grain, over this window, these two series moved together this much. The Epps effect is the finding that the claim degrades systematically as the grain gets finer. Two instruments that show strong correlation on daily bars show less on hourly bars, less again on minutes, and near nothing at seconds — not because the relationship weakens, but because the measurement stops being defined. Thomas Epps documented it in US equities in 1979; it has since been reproduced in FX, futures and crypto, and it is one of the most replicated findings in high-frequency finance.
The mechanism is mundane, which is what makes it dangerous to intuition. Two instruments almost never trade at the same instant. At coarse intervals that barely matters — a day contains thousands of prints from each side, and the interval averages over the asynchrony. At fine intervals it dominates: a one-minute bar may contain a print from one instrument and none from the other, so the return pair the formula compares is one real move against one interpolated flatline. Non-synchronous trading, discrete prices, and short-horizon lead-lag between venues all push the measured number toward zero. The literature’s corrective estimators — Hayashi–Yoshida most prominently — work precisely by comparing only genuinely overlapping observations, which is an admission of what the naive number was doing: measuring the sampling, not the relationship.
The practical consequence runs one way: fine-grain correlation readings understate, and fine-grain “decoupling” is usually the measurement dissolving rather than the relationship breaking. A pair that tracks tightly on the daily chart and “diverges” on the one-minute chart has, most of the time, done nothing of the kind — the observer has walked below the interval where co-movement is observable and mistaken the fog for an event. This is distinct from a genuine regime shift, where the relationship itself changes at a grain where it is well-measured; the two are separable by checking whether the coarse-grain correlation moved too.
It is measurable from data you can hold: compute the same pair’s correlation at a ladder of intervals — one minute, five, thirty, hourly, daily — over the same window, and the attenuation curve appears without any special access. Feed properties compound it: quote staleness and venue aggregation add artificial asynchrony on top of the real kind, so the same pair can dissolve at different grains on different feeds — which makes the curve, measured on your own venue’s data, a property of your data and not just of the market. Provenance matters for the same reason: what a file’s tick record actually sampled bounds what any correlation computed from it can mean.
Commonly confused with
Neighbouring concepts that get used interchangeably, and the distinction that actually separates them.
- Decoupling
Fine-grain decoupling is usually the measurement dissolving rather than the relationship breaking. A pair that tracks tightly on the daily chart and diverges on the one-minute chart has, most of the time, done nothing of the kind — the observer walked below the interval where co-movement is observable and mistook the fog for an event.
- Regime shift
A genuine shift changes the relationship at a grain where it is well measured. The Epps effect changes only the measurement. The two are separable by a single check: did the coarse-grain correlation move as well.
- Low correlation
A correlation number is a claim about a specific sampling interval, not a property of the pair. The same two instruments honestly carry different correlations at daily, hourly and one-minute grains, and none of those numbers is the true one.
- Quote staleness
Staleness adds artificial asynchrony on top of the real kind, so the same pair can dissolve at different grains on different feeds. That makes the attenuation curve a property of your data as well as of the market.
How to measure it in your own data
A definition you cannot test is a definition you have to take on trust. This is the shortest honest route from the concept to a number you computed yourself.
- Records you need
Both series over the same window, at the finest grain you hold, from a feed whose capture path you can describe. No special access is required — this is measurable from data you already have.
- What you compute
Compute the same pair's correlation at a ladder of intervals — one minute, five, thirty, hourly, daily — over the identical window. The attenuation curve appears on its own.
- What the answer tells you
The consequence runs one way: fine-grain readings understate. If the curve falls as the grain shrinks while the daily figure holds, you are looking at asynchrony rather than at a relationship ending. Corrective estimators — Hayashi-Yoshida most prominently — work by comparing only genuinely overlapping observations, which is an admission of what the naive number was doing: measuring the sampling rather than the relationship.
Questions and answers
My pair correlation collapsed on the one-minute chart. Has the relationship broken?
Almost certainly not. Two instruments rarely print at the same instant, and at fine intervals a bar may contain a move from one and nothing from the other — so the formula compares a real return against an interpolated flatline and reads the asynchrony as independence. Check the daily correlation over the same window before concluding anything.
Which sampling interval gives the true correlation?
None of them, in the sense the question implies. A correlation is a claim about a specific grain over a specific window, and the pair honestly carries different numbers at different grains. The useful output is the attenuation curve across intervals rather than a single figure.
How do I tell the Epps effect from a genuine change?
Check whether the coarse-grain correlation moved too. The Epps effect degrades the measurement as the interval shrinks while leaving the daily relationship intact; a real regime shift changes the relationship at grains where it is well measured. That one comparison separates them.
Does my data feed affect the result?
Yes, and more than most people expect. Staleness and venue aggregation add artificial asynchrony on top of the genuine kind, so the same pair can dissolve at different grains on different feeds. The curve you measure is a property of your data as well as of the market, which is why what the file actually sampled bounds what any correlation computed from it can mean.
Related terms
Derived from the links this entry makes and the entries that link back to it.
In the research
Epps Effect comes up in two research notes on this site.