Public FX Last-Look Latency Data and Deep-Tail Modeling
Summary
The document surveys the search for public measurements of electronic FX liquidity-provider response times, especially the far tail of last-look hold-time distributions during volatility spikes. It describes existing sources as incomplete: an older TCA study reports anonymized latency percentiles but only reaches relatively short delays, while venue disclosures offer typical or threshold figures. Academic work is mentioned as providing model structures without underlying response-time data.
The text asks whether public datasets or studies characterize extreme tail percentiles and proposes possible ways to model tails when observations are unavailable, including peaks-over-threshold methods with a generalized Pareto distribution, queueing models, and Hawkes processes for clustered congestion. It does not resolve whether suitable public data exist or recommend a validated model. The cited evidence is described at a high level, so the document is a research question and landscape overview rather than a reproducible empirical analysis.
Key ideas
- Public FX latency disclosures and TCA studies may not capture extremely delayed responses.
- The question focuses on response-time tails during high-volatility congestion.
- Peaks-over-threshold methods and generalized Pareto distributions are proposed for unobserved tail behavior.
- Queueing models and Hawkes processes are suggested for congestion and clustered delays.
- The document does not establish a definitive public dataset or validate a tail model.
Tags
Full text
# Are there any public datasets on FX venue/LP response-time (last-look hold-time) latency distributions? # Are there any public datasets on FX venue/LP response-time (last-look hold-time) latency distributions? I'm studying the latency of liquidity-provider responses in electronic FX (the time between sending a trade request to a venue/LP and receiving the fill or rejection). I'm particularly interested in the far tail of this distribution: the probability that a response arrives unusually late (say, beyond ~1 second), which appears to be driven by congestion during high-volatility periods (economic releases, other announcement-driven traffic spikes) rather than by normal hold-time variation.What I've already found: - The LMAX Exchange FX TCA white papers (2016 third-party-aggregator data, 7 LPs) publish anonymized execution-latency and hold-time distributions, including 50th/99th/99.9th percentile tail structure per venue. Useful, but it's 2016, aggregated, and the tail tops out around ~120 ms at the 99.9th percentile — it doesn't reach into the far tail I'm interested in. - Databento's microstructure notes and various venue disclosures (e.g. EBS ~12 ms average / 200 ms max threshold in 2022; Cboe FX threshold reductions toward 10 ms in 2025) give typical/last-look figures. - Academic work (Cartea et al. on last-look; the HSBC/Barzykin rejection-feedback paper) provides model structure but no underlying data. My questions: - Are there any publicly available datasets — even anonymized, aggregated, or academic — that report FX venue/LP response-time or last-look hold-time distributions, ideally with tail percentiles or measured during stressed/high-volatility periods? - Failing raw data, are there published TCA studies or venue disclosures that characterize the tail of the response-time distribution (beyond the 99.9th percentile), rather than just median/mean hold time? - When the empirical deep tail is effectively unobservable from public data, what are the accepted parametric approaches for modelling it — e.g. GPD / peaks-over-threshold, queueing-based congestion models, or self-exciting (Hawkes) processes to capture clustering during volatile periods? Any pointers to data sources, disclosure documents, or modelling references would be appreciated.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.