A median tick-to-trade that looked competitive was hiding a 99th percentile that decided the profitability of the book.
The firm had invested in colocation and a low-latency network stack, and its median tick-to-trade compared well with what it believed competitors achieved. Profitability on the market-making book nonetheless kept degrading, concentrated in fills that arrived immediately before adverse moves. The measurement in place reported an average, so the behaviour responsible for the losses was invisible.
We instrumented the full path — network transit, feed handling, decision, order construction, egress — and reported the distribution rather than the mean. The tail turned out to be dominated by allocation and serialisation in the feed handler, not by the model. The hot path was rewritten around preallocated buffers and a fixed message layout, with busy-poll networking and pinned cores to remove scheduler variance.
The median moved very little, which was the point. The 99th percentile fell by roughly two thirds and the distribution tightened, and the pattern of fills immediately preceding adverse moves reduced correspondingly. The firm now sizes its latency investment against a measured distribution rather than an average.
Fewer strategies reached production, and the ones that did behaved in live trading roughly as the research had said they would.
