Adverse selection in automated markets

Adverse selection is usually introduced as the cost of trading with a counterparty who knows something you do not. In a market full of human discretion, that framing is intuitive and roughly accurate. In an automated market it is intuitive and almost useless, because the counterparty is rarely a single informed participant. The counterparty is a population of quoting systems, each of which selects when to be present and when to be absent on the basis of signals you may or may not also be reading. The cost shows up not in any single fill, but in the conditional distribution of price moves that follow the fills you receive.

Aggregate fill-quality metrics are largely blind to this. A seventy-per-cent fill rate is silent on whether the thirty per cent that did not fill were the orders the desk most wanted; a slippage figure averaged over a day says nothing about whether the fills that closed at the touch were the ones immediately followed by an unfavourable move. The relevant joint distribution — fills cross-tabulated against the drift in the seconds after each fill — is the smallest measurement that captures the cost. Anything coarser will record adverse selection only as unattributed variance in a downstream P&L number, where it cannot be acted upon.

When that joint distribution is plotted, the shape of the asymmetry is more informative than the absolute level. A venue with symmetric counterparties produces a roughly symmetric drift distribution around each fill: the desk is sometimes filled before the market moves favourably, sometimes before it moves adversely, and the two cancel in expectation. A venue with informed makers produces a skewed distribution, with mass concentrated on the adverse side. The skew is the cost. It is also the cost that compounds most quietly — a few basis points per fill, absorbed into the noise of a daily report, until enough fills have accumulated for the statistical signal to dominate.

The most tractable auxiliary signal is the maker cancel-rate on the same instrument: how often, in a representative window, resting quotes are pulled in response to incoming flow before any match has occurred. The two distributions tend to move together. On instruments and hours where the resting quotes are largely static, fills look symmetric and the post-fill drift is approximately mean-zero. On instruments where resting quotes are pulled aggressively as flow arrives, the desk’s realised fill rate against its own best quotes degrades, and the fills that do occur are concentrated in the windows where the maker preferred not to cancel — which is, by selection, the windows the desk preferred not to be filled.

The conventional responses to a high adverse-selection regime are to widen the spread or reduce the order rate. Both are effective at reducing the rate at which the cost accrues, and both are silent about whether the desk should be quoting in that regime at all. The structural response is to treat adverse selection as a venue-level and hour-level parameter rather than a cost to be paid: to identify, with the same per-fill rigour as the rest of the lifecycle, the conditions under which the desk has the offsetting information — whether through its own signals, its routing topology, or its choice of instruments — and the conditions under which it does not. Quoting in regimes where the asymmetry runs against the desk is a position, not an execution detail. Recognising it as such is the prerequisite for declining it.

Adverse selection is, in the end, the cost that distinguishes venues, hours, and instruments where the desk has an edge from those where it is contributing one to someone else. The right unit of measurement is not a single fee number but a conditional distribution, computed at the order level and reviewed at the venue level. It is the measurement that, on a sufficient sample, will tell the desk where it should be present, where it should be absent, and where its existing analytics have been quietly attributing structural cost to noise.

← Back to Insights