Capital Markets / United Kingdom / 2023
Reducing signal-processing latency in a news-driven trading system
A proprietary trading firm needed millisecond-level reaction to high-frequency news feeds. We replaced the tokenisation stage of its sentiment pipeline with cache-resident probabilistic filters, cutting signal-processing latency by an order of magnitude.
Results
- 10×
- Reduction in signal-processing latency
Disciplines
- Low-latency systems
- Signal processing
- Performance engineering
Context
The firm traded on machine-readable news. Its edge depended on the interval between a headline arriving and an order reaching the venue, and the dominant cost in that interval was not the network — it was the firm’s own sentiment pipeline. Tokenising each message before classification put an unbounded, allocation-heavy stage directly on the critical path.
Approach
We began by instrumenting the existing path to establish where time was actually spent, rather than where the team believed it was spent. With the tokeniser confirmed as the bottleneck, we designed a classification stage that avoided tokenisation altogether: a set of Bloom filters sized to remain resident in L1 cache, evaluated against the raw byte stream.
The design was constrained as much by what it must not do as by what it must. No dynamic allocation on the hot path. No branching that the predictor could not learn. A false-positive rate low enough that downstream confirmation remained cheap, and a false-negative rate of zero for the term set that mattered.
Outcome
Signal-processing latency fell by an order of magnitude, moving the pipeline off the critical path and returning the firm’s response time to a regime set by its network and venue connectivity. The filter set is maintained by the firm’s own quantitative team; we designed it to be tuned without our involvement.