Last look, as a feature, exists for a reason. An LP streaming a continuous price into a network of takers is exposed to latency arbitrage — informed counterparties that read a public market move faster than the LP can refresh its own quote, then sweep the stale price. Last look gives the LP a small window — typically 5 to 50 milliseconds — to verify the quote against current market conditions before it accepts a fill. Used symmetrically and within a tight time bound, it is a defensible feature of a fair market.
Used asymmetrically, it is a transfer of expected value from the taker to the LP, paid in fills that fall through and in re-quotes against the taker's interest. The asymmetric version is statistically detectable, but only by counterparties that maintain the right instrumentation and the right tests.
This piece describes what asymmetric last look actually looks like in the fill tape, why it persists in a market where everyone claims symmetric behaviour, and how a counterparty can detect it from its own data without relying on the LP to self-report.
Symmetric last look, defined
Symmetric last look has three properties. First: the holding window is fixed and disclosed — typically a single configured value per LP, not adjusted dynamically. Second: the acceptance criterion is two-sided — if the market moves against the LP by more than the configured tolerance the fill is rejected, but if the market moves in favour of the LP the fill is accepted at the original quote. Third: the rejection rate, in steady-state conditions, is independent of which side of the trade is the taker.
In a market where all three properties hold, the LP recovers its latency-arbitrage premium without transferring expected value from the taker. The taker's reject rate is small, evenly distributed across winning and losing fills, and uncorrelated with the short-term direction of the market.
Asymmetric last look, in three statistical signatures
Asymmetric last look breaks each of those properties in one of three ways. Each break leaves a signature in the reject tape that an attentive counterparty can detect.
Signature one: directional rejection bias
On a symmetric LP, the reject rate on fills where the market moved against the LP in the holding window should equal the acceptance rate on fills where the market moved in favour of the LP. The two numbers should be within tight statistical bounds of each other, conditional on the absolute size of the move.
On an asymmetric LP, the reject rate is consistently higher when the LP would have lost on the fill than when it would have won. The asymmetry is detectable as a directional skew in the reject distribution, controlling for the magnitude of the price move during the hold. A counterparty with even a few thousand fills against the LP can run the test at p < 0.01 on a normal week.
Signature two: dynamic holding windows
On a symmetric LP, the hold is a fixed integer milliseconds. On an asymmetric LP, the hold is variable — sometimes 5 ms, sometimes 35 ms, sometimes more — and the longer holds cluster in conditions where the LP has higher uncertainty about its own quote. The signature is a multi-modal distribution of accept-time-from-quote, with a low-volatility mode at one value and a high-volatility mode at a substantially longer one.
Detecting dynamic holds requires the counterparty to timestamp both the quote and the accept message with sub-millisecond precision. Most counterparties cannot do this without help from the venue. A counterparty that runs its own colocated infrastructure and timestamps both sides of the FIX session has the data.
Signature three: post-reject reversion
On a symmetric LP, the residual market move in the second after a reject is unbiased — sometimes the market moves further against the original direction, sometimes it recovers. On an asymmetric LP, the post-reject market move is consistently against the taker — the rejection occurred precisely because the LP could profitably defer, and the residual move is the realised cost of the asymmetry.
Post-reject reversion is the most direct measure of the realised cost of last look to the taker. It does not require the counterparty to model the LP's internal decision process; it only requires a per-LP reject log with timestamps and the prevailing mid at +50 ms, +250 ms, and +1 s after the reject. The mean reversion is the cost.
Why asymmetric last look persists
If the asymmetry is statistically detectable, why has the market not priced it out? Three reasons.
First: most takers do not maintain the instrumentation to detect it. The required per-LP timestamping, reject taxonomy and post-reject mark-out infrastructure is rare outside the largest desks. Most counterparties see an aggregate reject rate, judge whether it is acceptable, and move on. The asymmetry is invisible without the underlying tape.
Second: the LPs that benefit from asymmetric last look are also LPs that streams competitively tight top-of-book. The taker that switches away on the basis of asymmetric reject behaviour incurs a visible cost in quoted spread that may not be offset by a less-visible saving in realised cost. The trade-off is real but rarely measured.
Third: the standard contractual language around last look is permissive. Most ISDAs and equivalent counterparty agreements describe last look in terms that allow for asymmetric implementation without it being technically a breach. Detection by the counterparty is therefore not the same as recourse from the counterparty.
What detection requires, operationally
A counterparty that wants to detect asymmetric last look on its own flow needs four data elements per fill attempt, per LP, persisted to a queryable store: (1) the quote price and timestamp at the moment of decision, (2) the side and notional of the attempt, (3) the LP's accept-or-reject response and the timestamp of that response, (4) the prevailing mid at three reference points after the response — typically +50 ms, +250 ms, +1 s.
With those four elements over a few thousand attempts per LP, the three statistical signatures above are computable in a few hours. The same instrumentation supports per-LP reject reporting back to the desk, which is the basis of the LP-level negotiation that actually moves the needle on realised cost.
How Drovix handles this
Every FIX session that Drovix maintains with an LP is timestamped at the wire at both the quote and the accept-or-reject boundary. Every fill attempt is logged with the full four-element record described above, and the per-LP reject taxonomy — including the three asymmetry signatures — is computed daily and exposed to the counterparty through TCA.
The counterparty sees, per LP: the symmetric-baseline reject rate, the directional skew (with a confidence interval), the hold-time distribution, and the post-reject mark-out averaged over the last 30 days. When an LP's signature drifts toward asymmetric behaviour, the routing weight on that LP drops automatically, and the counterparty's coverage desk is notified.
This is not a marketing claim about quoted spread. It is a description of what the instrumentation does when last look is, in fact, asymmetric — and how the counterparty avoids paying for it without having to detect it themselves.
The economics of the rejection bias compound through the broader routing decision. Queue-position economics at sub-millisecond granularity covers the routing side; the anatomy of effective spread covers where last look fits in the broader cost decomposition.
Analyst Desk
Drovix Research Desk
Institutional Research
Drovix Research Desk publishes institutional-grade analysis covering macro events, cross-asset correlations, and execution insights for professional market participants.