BL-EC-003: Adverse Selection
BIJECTIVE LABS
BL-EC-003: Adverse Selection
Status: RELEASED
Classification: PUBLIC
Revision: 1
Date: 2026-09-07
1. Why this is a separate axis
Adverse selection is usually filed under cost. In this record it has its own axis of death, and the reason is that it does not behave like a cost.
A fee is a constant. Slippage scales with size. Adverse selection scales with how attractive your quote is, which means it cannot be escaped by quoting better. The more competitive the price you show, the more of the informed flow you receive, and the informed flow is the flow that is about to be right.
That property inverts the usual optimisation. For every other cost term, improving your terms improves your outcome. For this one it does not.
2. What the universe quotes
The first question is whether the spread is wide enough to be worth earning. It is, on most of the venue.
A three-pass census of all 676 perpetual instruments [E-0051]:
| Statistic | Value |
|---|---|
| Median quoted spread | 5.26 bps |
| Mean quoted spread | 8.08 bps |
| 25th percentile | 2.40 bps |
| 75th percentile | 10.21 bps |
| 90th percentile | 18.54 bps |
| Instruments with two ticks of room or more | 62.1 pct |
| Instruments at one tick or less | 37.9 pct |
Against a maker round trip of 4.0 basis points, most of the universe looks profitable to quote on the fee schedule alone.
Two caveats belong on this table and both come from the artifact rather than from any narrative.
The maker fee was marked unverified at the time of measurement and reported as a grid rather than a confirmed schedule. The census establishes what the universe quotes, not what a specific operator nets.
Depth was gated on only 33 instruments. Spread without depth is a price nobody can transact meaningfully at, so the 62.1 percent figure describes quoted width rather than available liquidity.
3. What the fills cost
The second question is what happens after you are filled, and this is where the door closes.
Markout measures post-fill price movement in the direction that hurts the passive side. Measured on the control instrument across 9,352,722 trades [E-0002]:
| Horizon | Markout | Half-spread proxy |
|---|---|---|
| 60 s | -0.53 bps | 0.14 bps |
The adverse movement after a fill is roughly four times the half-spread being earned.
Across the candidate set, adverse selection exceeds the half-spread within ten seconds on five of six instruments. The control instrument measured negative on five days of five.
The single candidate that appeared to survive was rejected on concentration: 72.5 percent of its result came from a single day. A market-making result that lives in one day is not a market-making result.
4. Ten seconds is the number that matters
The measurement says the effect completes within ten seconds. That is the parameter a reader should carry away, because it is what makes this envelope-bound rather than universal.
Ten seconds is a long time in execution terms. A colocated participant with a cancellation path measured in microseconds has many opportunities to pull a quote inside that window. This envelope reacts on an order path of 5 to 6 milliseconds and a data path with a 2.6 second tail at the 99th percentile, which leaves considerably less room than the raw order latency suggests.
The verdict is therefore that adverse selection beat this operator's cancellation capability, not that it is unbeatable.
5. The same effect, measured from the other side
The most precise measurement of adverse selection in the archive was not taken on the venue where the market-making door closed. It was taken on the venue with identity-level data, and it decomposes the effect by who is doing the taking.
| Measurement | Value |
|---|---|
| Adverse selection difference, large against retail aggressors | 0.906 bps |
| CI95 | [0.757, 0.983] |
| Permutation p | 0.000 |
| Sample | 4.51M fills |
| Persistence | 10 to 30 s |
The two measurements agree on the mechanism and on the timescale. Adverse selection is real, it is driven by who is transacting, and it resolves inside tens of seconds.
They also agree on the economics, from opposite directions. On one venue the effect closes the market-making door. On the other, the effect is precise enough to be a tradeable signal in principle and too small to clear a round trip in practice, at 0.906 basis points against 3.0 at the cheapest available channel.
6. What this means for a reader
Three conditions would have to hold for passive market making to work on this venue set, and the record can price the first two.
A cancellation path fast enough to act inside ten seconds. Not achievable
from a non-colocated cloud instance under load, on the evidence in BL-EC-002.
A fee schedule where the earned spread survives the markout. At a median quoted spread of 5.26 basis points and measured adverse selection of roughly four times the half-spread on the control, the schedule would have to improve substantially rather than marginally.
Instrument selection that is not concentration. The one apparently viable candidate was 72.5 percent one day. Any selection procedure has to survive a concentration check, and this one did not.
The first two are envelope parameters. The third is a discipline.
7. What was not measured
Adverse selection by time of day, by regime, or conditioned on any of the volatility signals the programme validated elsewhere. The measurement is pooled.
That is a real gap and it is a plausible route. If adverse selection concentrates in identifiable conditions, a quoting strategy that stands down in those conditions is a different proposition from one that quotes continuously. The programme did not test it, and the record says so rather than implying the question was settled.
8. Cross-references
BL-D-005 on why this is a separate axis. BL-EC-002 for the latency the
ten-second window is measured against. BL-P2-001 door 3. BL-P2-003 sections
2.4 and 2.6. BL-SD-026, BL-SD-027, BL-SD-046. Evidence E-0002, E-0051.
Copyright (c) 2026 Bijective Labs. All rights reserved.