The Half-Life of Information in FX Orders

A trade carries information for as long as the market has not yet priced in what the trader was reacting to. After that, the trade is news only to the people who do not know it happened. The interval between those two states — between the moment of execution and the moment the information is fully impounded — is the half-life of the information embedded in the order.

Half-life is a useful frame because it is empirical, asset-specific, and tractable. In the most-liquid FX pairs and equity-index futures, the half-life of an informed order is on the order of seconds to a minute. In a thinly-traded EM cross or a single-name CFD on a small-cap, it can be many minutes. The number matters because it determines how much of the original signal the desk can still capture on a child clip if it executes ten seconds later, sixty seconds later, five minutes later.

It also matters because the LP, on the other side of the trade, is running its own half-life estimate — usually informally, sometimes with explicit instrumentation — to decide how much adverse-selection premium to charge a given counterparty. The half-life governs the rate at which the LP's memory of bad flow fades, and therefore the rate at which a desk that has been mis-identified as informed can earn back tighter pricing.

Measuring half-life directly: the mark-out curve

The half-life of information in a desk's flow is computable directly from its own fill tape. For every fill, compute the signed price move from the execution price to the prevailing mid at a sequence of horizons: +1 s, +5 s, +30 s, +2 min, +10 min, +1 h. Sign the move so that a favourable move (mid moved with the side of the trade) is positive and an unfavourable move is negative.

Average across all fills in the desk's history. The resulting curve — average signed mark-out as a function of horizon — is the empirical fingerprint of how much information the desk's flow carried at the moment of execution. A flat curve at zero means the desk's flow was uninformed; the LP made its quoted spread cleanly. A curve that climbs above zero and plateaus means the flow was informed and the LP was paying the desk a positive expected value at execution. A curve that climbs and then reverts means the flow was momentum-following: informed in the short term, decaying back to zero in the longer term.

The half-life of the curve — the horizon at which the curve reaches half of its eventual plateau — is the desk's identifiable signature to its counterparties. On most institutional FX desks, the curve plateaus within 30 seconds. On desks with a slower discretionary signal, it can take 5 to 10 minutes.

Why the LP cares about half-life

The LP cannot read the desk's signal. It can only observe its own realised mark-out on the desk's flow. If the LP's mark-out curve against the desk is consistently positive at short horizons and plateaus at, say, 0.5 pips after 30 seconds, the LP is paying the desk 0.5 pips per fill in expected value. The LP recoups that by widening the adverse-selection premium it charges the desk on subsequent quotes.

The LP's widening of the premium is not instantaneous. It is a moving average over the desk's recent flow, weighted by the LP's confidence in its estimate. The rate of decay of that moving average — how quickly the LP "forgets" a streak of informed fills once the desk's flow normalises — is itself a kind of half-life, and it is the variable the desk has the most ability to influence.

How long does it take to earn back tighter pricing?

Empirically, across a broad sample of institutional FX desks reporting through standardised TCA frameworks, the half-life of an LP's adverse-selection memory is on the order of one to three weeks of typical flow. A desk that has been flagged as informed — by, for instance, a single news-trading session — pays a wider spread for one to three weeks before the LP's estimate normalises, even if the desk's intervening flow is uninformed.

This number is consequential. A desk that pays an extra 0.3 pips of adverse-selection premium for two weeks on EUR/USD flow of $50M per day is paying roughly $15,000 in unnecessary cost. The same desk, if it understood the half-life, would either (a) avoid news-trading on the venue where it needs tight steady-state pricing, or (b) carry a second venue for news exposure, isolating the LP-memory effect.

Routing as memory management

Once a desk treats LP memory as a managed variable, routing becomes a memory-management problem. The desk allocates its uninformed flow — passive hedges, scheduled rebalances, low-information trades — to LPs where it wants to maintain a tight steady-state quote. It allocates its informed flow — discretionary trades against a short-term signal — to LPs that price the adverse-selection premium more aggressively at the entry but recover their memory faster, or to venues where the desk does not need a long-term relationship.

This is not a market-microstructure abstraction. It is the same logic that a credit-card user follows in deciding which card to use for which purchase — minus the marketing, plus a non-trivial amount of math.

What the counterparty can do for the desk

A counterparty that aggregates across multiple LPs has visibility into the per-LP mark-out curve that no single LP has. It can route a desk's flow to the LP whose memory is most favourable in the current moment, conditional on the asset, the size, and the desk's recent signature. It can also report back to the desk what its current signature looks like across its LP panel — which lets the desk decide consciously whether it wants to flag itself as informed on a particular venue at a particular time.

Drovix exposes per-LP mark-out curves to every counterparty on its institutional portal, refreshed daily. The desk sees, per LP, the average signed mark-out at five horizons and the implied current adverse-selection premium. The routing engine consumes the same curves and rebalances flow on the basis of them. The decomposition of the adverse-selection premium into its actual constituents is covered in adverse selection premiums and how they're priced; the routing layer that consumes the half-life signal sits next to queue-position economics.

The half-life of the relationship itself

There is a meta-point in this piece that institutional desks should not miss. Every commercial relationship between a desk and a counterparty has its own half-life — the rate at which a single bad month of execution decays from the desk's memory of the counterparty, and from the counterparty's memory of the desk. A desk that switches counterparties at the first bad month never accumulates the steady-state pricing that long memory permits. A desk that never switches accumulates entrenched costs that no instrumentation can recover.

The half-life of a relationship is, in practice, around six to nine months — long enough for steady-state pricing to consolidate, short enough that a desk that is being systematically mistreated does not pay for it indefinitely. The counterparties that win institutional business at this horizon are the ones that publish the data — per-LP mark-out, per-LP reject behaviour, per-LP routing weight — that lets the desk verify on a weekly basis that the steady state is still earned.

Drovix runs this exposure as a deliberate product feature. If the steady-state pricing across the panel drifts, the counterparty sees it before it compounds into a quarterly cost.