Skip to content
Nexus Prop LLC

Methodology guide

How to measure latency, and how to read someone else's numbers.

Most published trading-platform latency figures are unfalsifiable. This is how to measure honestly, and how to interrogate a number someone else is quoting at you.
By Ethan Warmuskerken, Founder, Nexus Prop LLCLast reviewed

Key takeaways

  • A latency figure without a path, sample size, date and percentile distribution is marketing, not measurement.
  • Averages hide exactly the behaviour that matters; p99 and p99.9 are where traders actually feel a platform.
  • Ask what is excluded from the measurement — that is usually where the interesting latency lives.

Define the path before the number

'Latency' is meaningless without endpoints. Order acknowledgement from the edge? Matching engine internal processing? Trader click to visible fill? These can differ by orders of magnitude, and vendors naturally quote the flattering one.

A credible figure states both endpoints explicitly, and ideally reports several paths so the reader can see where time is actually spent.

Report distributions, not averages

An average latency of 5ms is compatible with a meaningful share of requests taking 500ms. Traders do not experience the average; they experience the bad tail, and they experience it at exactly the busiest moments.

Report p50, p95, p99 and p99.9 together with the sample size and the observation window. If a vendor reports only a single number, ask which percentile it is — the answer is often 'the average', and sometimes 'the best'.

What a credible latency claim states

ElementWhy it matters
Measured path (both endpoints)Determines what is actually being timed
Percentiles p50/p95/p99/p99.9Averages conceal the tail traders feel
Sample size and windowA good minute is not a claim about a platform
Geography and network pathClient-observed latency depends heavily on both
Load conditionsQuiet-period figures do not describe a market open
Date measuredInfrastructure changes; a year-old figure describes a different system
What is excludedQueueing and backlog are often quietly outside the measurement

Backlog and failures belong in the report

A system can post excellent latency while dropping or deferring work. Any honest report includes the failure rate and any backlog during the measurement window, because a fast path that sheds load is not fast — it is incomplete.

This is also why a live 'latency ticker' on a marketing site is usually decorative. Unless it states path, sample, percentiles and methodology alongside the number, it is an animation. There is no evidence that publishing one improves search performance, and a number nobody can interpret does not build trust with technical buyers.

Ask us for numbers with methodology.

We would rather give you a figure you can interrogate than one you have to trust.