BL-EC-002: Latency, and What It Places In Reach
BIJECTIVE LABS
BL-EC-002: Latency, and What It Places In Reach
Status: RELEASED
Classification: PUBLIC
Revision: 1
Date: 2026-09-07
1. The measured profile
Taken from a cloud instance in Singapore against the Singapore matching engine, on a persistent connection.
| Path segment | Measured |
|---|---|
| Network TCP round trip to venue | 3 to 4 ms |
| Order round trip, persistent connection | 5 to 6 ms |
| Order round trip, cold TLS handshake | approximately 93 ms |
| Internal order queue hop | 0.094 ms median, 0.267 ms p99 |
| Market data freshness over websocket | approximately 80 ms median, 1.4 s p90, 2.6 s p99 |
| End-to-end event-driven reaction floor | 10 to 20 ms |
| Production polling configuration in use | 250 to 260 ms |
On the second venue, block cadence is approximately 70 ms and the round trip from Singapore is approximately 96 ms, without colocation at any point.
2. Two numbers that are usually conflated
The order round trip is 5 to 6 milliseconds. Market data freshness is approximately 80 milliseconds at the median and 2.6 seconds at the 99th percentile.
The second is fifteen times the first at the median and three orders of magnitude worse in the tail. That asymmetry is the practical constraint, and it is not what most latency discussions focus on.
Being able to send an order in six milliseconds is worth nothing if the price that triggered the decision was eighty milliseconds old when it arrived. The binding number for a reactive strategy is the data path, not the order path, and it is the one that is harder to improve without colocation.
The 2.6 second tail matters more than the median for a different reason. A strategy that assumes fresh data will occasionally act on a picture two and a half seconds stale, and it will do so precisely during the volatile periods when staleness is most expensive.
3. The gap between the floor and the configuration
The event-driven reaction floor was measured at 10 to 20 milliseconds. The configuration actually running polled at 250 to 260.
That gap is deliberate and is worth explaining, because it looks like a defect.
Event-driven paths were built and used where a hypothesis required them. The tick-exact dislocation work ran on one. Polling was used elsewhere because the polling path is simpler to reason about, easier to reproduce in a backtest, and because the additional 240 milliseconds is immaterial to every hypothesis tested at four-hour horizons and beyond.
R-13 closes directional hypotheses below four hours on cost. Once that rule is
in force, sub-100-millisecond reaction has nothing left to buy, and the
engineering effort to close the gap would have produced no research.
4. What this profile excludes, by design
Latency arbitrage and queue-position contests against colocated participants were declared out of scope at layer zero rather than attempted and lost.
The reasoning is arithmetic. A participant colocated at the venue operates below one millisecond. This envelope operates at five to six on the order path and eighty on the data path. In a contest where the winner is determined by ordering, a fifty-fold disadvantage is not closable by better code.
BL-P2-003 section 2.5 records the speed class as excluded rather than
falsified, and the distinction is deliberate. Nothing was measured, so nothing
was concluded about whether speed edges exist on these venues. They are simply
not available from this position.
Recording an untested class as untested is a small discipline that prevents a larger error later, which is a reader concluding from this record that speed does not pay.
5. Where latency did decide a verdict
Two rows in the Kill Ledger turn on the profile above rather than on cost or on the idea.
Passive market making. Adverse selection exceeds the half-spread being
earned within ten seconds on five of six candidate instruments. Ten seconds is a
long time by execution standards and a short one by cancellation standards from a
non-colocated position. The row is classified ENVELOPE-BOUND on queue position
and reaction speed, and BL-EC-003 covers it.
Oracle-tether premium. The dislocation reverts with a median half-life of 3.9 seconds. Arriving at 96 milliseconds without colocation, and crossing rather than quoting, the strategy lost 2.56 basis points gross before any fee was applied. The reversion is real, fast, and not reachable from this position.
Both of these are honest latency findings, and both are recoverable by an operator with a shorter path.
6. What a shorter path would change
| Row | Currently blocked by | What a shorter path gives |
|---|---|---|
| D3, market making | Cancellation speed against informed flow | Time to pull a quote inside the ten-second window |
| S3, premium reversion | 96 ms against a 3.9 s half-life | A larger share of the reversion captured |
| Speed class | Never opened | Nothing this record can predict |
The third row is the honest one. This programme has no measurements of what a colocated participant would find on these venues, and the record does not speculate.
7. Cross-references
BL-FM-002 section 4 declares the parameter. BL-EC-003 for adverse selection.
BL-P2-003 section 2.5 for the speed class. BL-SD-024 and BL-SD-046.
Copyright (c) 2026 Bijective Labs. All rights reserved.