Adverse Selection Premiums and How They're Actually Quoted

When an LP quotes a price, it knows that some fraction of the counterparties on the other side of that price know something it doesn't. The LP cannot identify which counterparty falls into that informed minority. So it does the next best thing: it prices a small premium into every quote — the adverse-selection premium — that compensates it, on average, for the expected value transferred to the informed minority.

Everyone pays the premium. The informed minority recovers it by carrying information advantage. The uninformed majority pays it as a tax on participation. The size of the tax, the conditions under which it widens or compresses, and the variables that move it — these are the most material parameters in institutional execution and the least transparently disclosed.

This piece is a structural look at how adverse-selection premiums are actually computed by LPs in 2026, where the LP's flexibility sits, and which counterparty behaviours move the premium in the desk's favour.

The structural decomposition of the premium

From the LP's side, the adverse-selection premium has three structural components. The first is the unconditional baseline: the average expected value transferred to the LP's counterparties under the LP's current mix of takers, averaged across all conditions. The second is the conditional adjustment: how much wider or tighter the premium should be in the current market state — volatility regime, time of day, proximity to a scheduled event, recent realised toxicity. The third is the counterparty-specific overlay: the LP's estimate of the present desk's recent flow signature.

Sophisticated LPs run all three layers continuously. The baseline updates monthly. The conditional adjustment updates intraday. The counterparty overlay updates per-fill, with a configured smoothing constant — the same memory half-life discussed in the half-life of information piece. The combined premium quoted to a desk at a given moment is the additive composition of the three layers.

Less sophisticated LPs run only the first two layers. They quote a counterparty-blind spread and accept that informed flow will tax them; they recoup the tax through a wider baseline. The result is a market that is uniformly more expensive for everyone — the uninformed pay too much, the informed pay a fair price, the LP runs slightly looser margins than it could.

Where the desk has leverage

Of the three components, the only one a desk can move in its own favour is the counterparty-specific overlay. The baseline and conditional layers are properties of the LP's broader business and the prevailing market state. The overlay is the LP's estimate of this specific desk.

The desk moves the overlay by managing the LP's mark-out curve against its flow. Concretely: the desk concentrates its lowest-information flow on the LPs where it most needs tight steady-state pricing, and routes its highest-information flow either to venues that price adverse selection per-fill (more expensive on the day, but with no memory effect on subsequent days), or to anonymous venues where the LP cannot attribute the flow to a specific desk in the first place.

This is a two-dimensional optimisation: the desk pays a slightly higher per-fill premium on a subset of trades in order to keep its average mark-out on the remaining flow as close to zero as possible. The objective function — total realised execution cost — is minimised at an interior solution that is not obvious without the empirical mark-out curves for each LP.

The conditional layer: when premiums widen

Premiums widen, predictably, around scheduled news events. They widen also around month-end FX fixings, central-bank communication days, and treasury auction settlements. The widening is real and unavoidable — the LP's uncertainty about its own quote in those windows genuinely is higher — but the magnitude varies enormously across LPs.

A counterparty that aggregates across multiple LPs sees the cross-section of how each LP prices the same conditional event. Some LPs widen on news by 40%, others by 200%. The aggregated pricing engine routes around the latter to the former in the relevant windows, capturing the LP-specific overshoots as savings.

Premiums also widen around toxicity events the desk did not generate but is associated with. If the LP's recent realised P&L against a peer group of similar desks is poor, the LP may push the conditional layer wider for the whole peer group — even desks that have not personally generated toxic flow. This is a form of guilt-by-association that the desk can only escape by maintaining a long enough run of clean mark-out to differentiate itself within the peer group.

The counterparty-specific overlay: how the LP estimates the desk

The LP's estimate of the desk's flow signature is conceptually simple: it is the time-weighted average mark-out on the desk's recent fills, normalised by the average mark-out across the LP's broader book. The implementation is more involved — the LP has to control for the asset, the size, the time-of-day, the prevailing volatility — but the core quantity is well-defined.

Once the LP has the estimate, the overlay is a smooth function: an estimated signature of zero implies a small or zero overlay, an estimated signature of one pip implies a meaningful overlay on top of the conditional layer. The function's steepness — how aggressively the LP adjusts the overlay in response to a given signature — varies enormously across LPs, and is one of the few variables that distinguishes LPs commercially.

LPs that adjust aggressively recoup their losses quickly when they price a desk wrong, at the cost of giving up uninformed flow when the desk's signature is briefly noisy. LPs that adjust slowly absorb more losses when they price a desk wrong, at the benefit of maintaining tighter steady-state spreads. There is no single right answer; different LPs serve different niches.

What a counterparty's TCA should expose

A counterparty whose TCA does not break down realised cost into baseline-plus-conditional-plus-overlay is leaving the desk to wonder which of the three layers is driving its costs in a given month. A counterparty whose TCA does break it down lets the desk run the two-dimensional optimisation described above with actual data.

The breakdown requires the counterparty to maintain (a) a per-LP unconditional baseline reference, (b) a per-LP conditional adjustment as a function of regime, and (c) a per-LP per-desk overlay estimate from the per-LP mark-out curve. The first two are reproducible without proprietary access; the third requires the counterparty to maintain its own mark-out instrumentation against every connected LP, on every desk's flow.

Drovix maintains all three layers in production and exposes them per-LP through the institutional portal. The desk sees, for each LP, the current baseline, the conditional adjustment for the prevailing regime, and the desk's own current overlay estimate. The routing engine consumes the same decomposition to make per-clip routing decisions.

The relationship between LP overlay and counterparty memory is covered in the half-life of information in FX orders; the relationship between conditional layer and venue selection is in queue-position economics.