The Capacity Curve of Aggregated Liquidity: When One More LP Hurts

The intuition that more liquidity providers always make a counterparty's pricing better is wrong. The number of LPs on a panel has a capacity curve: pricing improves with the first few additions, plateaus through the middle, and degrades — measurably — past a particular point. The shape of the curve, and the specific point at which one more LP starts hurting, is one of the underrated engineering questions in institutional execution.

This piece is an account of the capacity-curve work Drovix has run over the last eighteen months on its connected LP panel, what the empirical curve actually looks like on a major institutional FX pair, and the specific mechanisms by which a marginal LP can produce a marginal negative — including the conditions under which a desk should ask its counterparty to drop an LP rather than add one.

The intuitive case for more LPs (and why it has limits)

With one LP, the desk pays the LP's quoted spread plus the LP's adverse-selection premium against the desk's flow signature. With two LPs, the desk routes the harder-to-fill clips to the cheaper venue and lets the easier clips fill on the better-quoted venue, and the average realised cost falls. With three LPs, the routing engine has more flexibility to allocate flow conditional on size, urgency and recent LP behaviour, and the average cost falls again.

This logic continues. With five LPs the routing has access to four-way price competition on every clip, with three of the five LPs being secondary candidates for the worst-case fill. By seven or eight LPs, the marginal improvement per added LP is small but still positive on a typical liquid major. By twelve LPs on the same major, the marginal improvement has reached zero. Past fifteen, on average, it has gone negative.

Why the curve eventually turns negative

Three mechanisms drive the negative slope at the high end of the curve.

First: the cost of message-rate management. Each connected LP delivers a continuous market-data stream. The aggregation engine must consume, normalise and merge those streams at the throughput rate of the slowest critical path. Past a certain number of LPs, the aggregation latency per market-data tick rises, the engine's decision time on each clip slightly extends, and the queue-position premium starts to leak. The leak is small per clip but consistent.

Second: the cost of LP-specific calibration. The routing engine maintains per-LP behavioural models: reject distributions, mark-out curves, recent fill quality, conditional adjustments per regime. Each LP requires statistical power to maintain its model with low variance. The desk's total flow is finite; spreading it across more LPs reduces the flow per LP, increases the variance of each LP's model, and degrades the routing engine's per-clip decision quality.

Third: the cost of LP toxicity correlation. Marginal LPs added at the long tail of the panel are often LPs whose stress behaviour correlates with each other — typically smaller specialist non-bank market makers whose risk capacity scales together in a stress. Adding them in normal times produces marginal tightening of top-of-book; in a stress, they all withdraw together, leaving the panel suddenly thinner than expected. The realised cost of the correlation hits in exactly the conditions the desk needed the depth most.

The empirical curve on a major pair

On EUR/USD, across the last 18 months of Drovix's panel performance, the empirical capacity curve as a function of connected LP count is roughly: average realised spread improves by ~0.25 pips from one LP to four, by a further ~0.10 pips from four to seven, by ~0.04 pips from seven to ten, and is essentially flat thereafter. Past twelve LPs, the curve begins to slope upward at roughly 0.01 pips per added LP, attributable to the three mechanisms above.

On thinner pairs — say USD/MXN or AUD/CAD — the curve plateaus earlier and turns negative sooner. The effective panel for a typical EM cross is six to eight LPs; adding more produces no measurable improvement and adds calibration noise.

On a major equity index CFD, the curve is shaped similarly but with a different inflection point — six to ten LPs is typically the productive range, with the optimal panel size depending heavily on the LPs' inventory horizons and overnight risk policies.

When a desk should ask the counterparty to drop an LP

The capacity curve implies that an actively-managed panel has churn. Adding an LP that improves the panel's pricing is the obvious half. Dropping an LP that no longer earns its keep is the half that almost nobody does.

A desk should consider asking its counterparty to drop an LP from its share of the panel when (a) the LP's realised mark-out against the desk's flow has drifted negative for more than two months, (b) the LP's reject rate has risen materially or asymmetric-last-look signatures have emerged, (c) the LP's conditional widening on news events has grown disproportionate to its peers, or (d) the LP's correlation with another LP on the panel — measured by joint behaviour in stress windows — has risen above an explicit threshold.

None of these conditions imply that the LP is doing anything wrong. They imply that, for this desk, the LP is no longer at the productive part of the desk's capacity curve. The counterparty should be able to drop the LP from the desk's panel without disrupting the desk's overall execution; that is a feature of a real aggregation layer.

How Drovix manages the panel

Drovix maintains a curated panel of bank and non-bank LPs, with documented per-LP performance against the metrics described above, reviewed quarterly. The curation includes both additions and removals; an LP whose performance has drifted out of the productive range is removed rather than carried as a marketing add-on.

The desk's per-counterparty panel is configurable within the curated set: a specific counterparty can request a narrower panel, exclude specific LPs, or request a panel weighted toward certain LP types. Each configuration is reflected in the counterparty's TCA, with the realised effects of the panel choice attributable per-LP.

The point is not that more LPs is always better. The point is that the optimal panel is a finite set, that it changes over time, and that the counterparty's job is to make those changes on behalf of the desk without churning the desk's connectivity or operational state. The capacity curve is the empirical justification for treating panel management as a managed product rather than a fixed feature.

The structural implication of capacity-curve thinking for diligence on a new counterparty is in counterparty concentration risk, and the implication for the routing engine itself sits next to queue-position economics.