How Fast Order Flow Alpha Decays

A signal like queue imbalance or trade aggressor imbalance predicts price direction — but only for a while. Measure the correlation between the signal and subsequent price changes at increasing horizons, and it starts strong immediately after the signal is observed and steadily fades toward zero. How fast that fade happens, the signal's decay horizon, is one of the most practically important numbers in any order-flow strategy, because it determines what kind of trading infrastructure and holding period the signal can even be used with.

The idea: a scent trail going cold

A tracking dog following a scent trail can follow it confidently minutes after it was laid, less confidently after an hour, and not at all after a heavy rain has washed it away. The scent's usefulness has a horizon, and a dog that waits too long before starting to track gets nothing. Order-flow signals behave the same way: the information in a burst of buy-heavy trades or a queue imbalance is real but fades — new order flow arrives, prices adjust, and whatever edge the original signal contained gets absorbed into the price or overwritten by newer information. A strategy has to act within that horizon, or the signal is worthless by the time it does.

Measuring decay directly

Decay is measured by computing the correlation, or the average subsequent price move, between a signal observed at time 0 and the price change over horizons of increasing length:

$$\text{Predictive power}(h) = \text{Corr}\left(\text{Signal}_0,\; P_{h} - P_0\right) ,$$

for horizons $h$ = 1 second, 5 seconds, 30 seconds, 1 minute, 5 minutes, and so on. In plain English: this asks "if I knew the signal's reading right now, how much would that have told me about the price move over the next $h$ seconds," repeated at increasing values of $h$ until the answer settles at essentially zero.

Worked example: a decay curve for queue imbalance

A researcher measures the correlation between top-of-book queue imbalance and subsequent price change at several horizons for a liquid large-cap stock:

Horizon	Correlation
100 ms	0.18
1 second	0.11
5 seconds	0.05
30 seconds	0.01
5 minutes	≈ 0.00

The signal has real predictive power at sub-second and single-digit-second horizons, is nearly gone by 30 seconds, and is indistinguishable from noise by five minutes. This tells the researcher immediately that queue imbalance is a signal for market-making quote adjustments and sub-second execution timing, not for a strategy that holds positions for minutes — trying to trade this signal with a system that takes 500 milliseconds to react has already burned off roughly half its edge before acting on it.

A decay curve has the opposite shape to the convergence plot above — instead of an estimate settling in on a stable truth as more data arrives, predictive correlation starts elevated and settles down toward zero as the horizon lengthens — but the same idea of "watching a quantity stabilize" applies to reading either chart correctly.

What this means in practice

Decay horizon is what separates strategies structurally: signals with millisecond-to-second decay are only usable by infrastructure fast enough to act within that window (co-located market makers, HFT desks), while signals with minutes-to-hours decay (order-flow aggregated at coarser intervals, or metaorder detection) are usable by ordinary execution algorithms and even discretionary traders. Measuring a signal's own decay curve, rather than assuming a textbook horizon, is a required first step before building any strategy on it — a signal validated at one horizon and deployed at another can look profitable in a backtest that doesn't account for realistic reaction latency and be worthless live.