Market microstructure programme
Our own programme — a market measured to the millisecond, and a result we published against ourselves
Sector: Quantitative research · Year: 2026
Summary
A high-frequency measurement programme built to settle one question properly: when a short-dated market and the feed that settles it disagree, is the gap an opportunity, or an artefact of how fast we are able to see? Answering it honestly required instrumentation an order of magnitude more careful than the trade it was evaluating. What we describe here is the measurement. The strategy work it serves is not published.
Challenge
A price difference on a screen is not an opportunity. Between a market and the feed that settles it sit transport delay, aggregation lag, book withdrawal and clock skew — and every one of those can manufacture a gap that looks exactly like edge and cannot be captured. The distinction cannot be drawn from historical bars. It needs the market recorded as it actually arrived: stamped on receipt, at full depth, continuously, for weeks.
Approach
- Recorded multiple venues in parallel at full order-book depth, stamping every record on receipt to the microsecond, into partitioned columnar storage designed so a month of tape can be scored without ever being loaded whole.
- Pre-registered the decision gate before the data existed — six criteria, thresholds frozen in writing, and the correction for testing three assets simultaneously fixed in advance at a 98.33% per-asset bound rather than chosen afterwards.
- Audited the tape against itself instead of trusting the recorder: exchange sequence contiguity, book invariants, settlement reconstruction against the on-chain outcome, and per-feed latency measured against each feed’s own publication clock.
- Split calibration, test and confirmation into tiers, holding back a live tape that did not exist when the analysis was written — so it could not be fitted to, even by accident.
- Instrumented the failure modes that counterfeit success: a frozen feed still emitting confident prices, a placeholder record that consumes a sequence number but represents no fill, and an accuracy measure that scores a deliberate refusal to price as a wrong answer.
Outcome
Continuous capture across three assets for twenty-five days: 6,445,779 ticks at 99.5% coverage, 21,269 settlements reconstructed with zero mismatches, and a separate venue tape audited at 2,534,160 of 2,534,160 trade identifiers present — zero loss, and no sequence break across 2.17 million book updates. The measurement then returned a decisive answer. The settlement feed reaches us 1,627ms after its own timestamp while the execution feed arrives in 152ms on the same clock, against a median opportunity that lives 1,000ms. The gap is real, and it is structurally uncapturable — a property of the relay, not of the sample, so no amount of further recording would change it. The programme closed the strategy on its own pre-registered criteria, on schedule, having committed no capital at any stage. The instrumentation, the audit battery and the documented null are what remain, and they are the point.
Stack
Python, asyncio, WebSockets, Apache Arrow, Parquet, NumPy, pandas
Published papers
- Beaten by the Relay — The opportunity lasted a second. The feed that defined it arrived 1.6 seconds late. A structural result, not a sample-size one. 7 pages, PDF.