Equity Trading: Performance, Latency & Throughput

Phil Penhaligan

Head of Equity Product Technology (LSE and Turquoise) at London Stock Exchange

Low Latency: What are we measuring?

• 3 distinct attributes

  • Measured from just outside LSE/TQ firewalls

  • Excludes transmission time to/from external client/venue

  • Excludes latency of client/venues's own systems(s)

  • Excludes network packet retransmissions due to slow consumers

  1. Order latency

     • Time from Order receipt to order acknowledgment

     • Commonly referred to as Order:Ack

  2. Market Data latency

     • Time from Order acknowledgement to public broadcast Commonly referred to as Ack:Tick

  3. Reference price latency (Turquoise Dark Book only)

     • Time from Tick receipt to matching engine usage

• 2 key metrics

  • Average

    • Mathematical mean

  • Consistency

    • Max 99.9 percentile

      • Discard the worst 0.1% of orders and then take the maximum

• Production Systems

  • Continuous real-time capture

  • Historical data stored in database for statistical analysis and reporting

• QA Systems

  • New software releases evaluated during technical tests

  • To confirm/verify expected latency improvements

  • To detect unexpected latency regressions

Low Latency: How Fast Is Fast Enough?

• Customer feedback

  • LSE/TQ averages are already fast enough - better Consistency is most important!

• Our Focus

  • Since the initial deployment of Millennium (Oct 2009 Turquoise and Feb 2010 LSE)

    • We have improved order:ack average by 30% & consistency by 90%

    • Multicast ack: tick has improved by 50%

  • We continue to focus on improving consistency (see later slides for details)

• My Observations

  • Since migration to Millennium IT, the exchange latency is an order of magnitude better than latency within client applications + transmission to/from our venues

    • Therefore for the majority of clients who are not co-located, further improvements to latency will have little impact

    • E.g. a 10% improvement to exchange latency = client 1% improvement

  • Co-location customers whose algorithms interact exclusively with LSE/TQ markets have the most to gain from further reductions in latency

  • Our latency is a tiny fraction (<0.0005%) of human reaction times (typically <200mS)

Low Latency: Who really needs it?

• Clients/Participants

  Category	Least Sensitive	Most Sensitive
  Manual Trading	Brokers, 'High Touch' Traders.	Arbitrage Traders, Portfolio Traders.
  Automated Trading	Retail Service Providers (RSP), DMA Clients.	Market Makers.
  Algorithmic engines & Smart Order Routers.	Remote Algorithms e.g. VWAP, POV, IS typically driven more by historical data curves than real-time ticks.	HFTs, Best Execution SORs, Co-lo Algorithms (e.g. arbitrage, momentum).

• The Venue/Exchange

  • Venues with the lowest latency Market Data are at the front of the queue when BBO updates occur

  • Quality of dark book trades depends on ref price latency

  • Builds strong technology reputation

  • Generally good for Sales & marketing

Low Latency: What's the next best thing?

• Consistency

  • Fewer outliers, and outliers closer to the average rather than simply seeking ever lower average latency is the most important feature for any trading strategy whether manual or automated.

  • If your system cannot compete on latency average, you can always make improvements by making it more consistent.

• Resilience

  • Closely tied to consistency is resilience, since an unresilient system will never be able to provide consistent results!!

Low Latency: Technology

Performance & Capacity Management - KPIs

Performance & Capacity Management

• Each system is technically tested with every software release to

  • Prove KPI levels

  • Reconfirm/prove behaviour and known breaking points

• Preliminary tests take place on pre-production

• At least one test cycle takes place on the actual production hardware on a weekend

• Golden Rules

  1. A client message will always get a valid response (ack or nack)

  2. Any component that has been taken so far above the KPI level that it fails must do so gracefully, and the system must continue to obey golden rule 1.

• Because we know the system/component limitations & bottlenecks we can manage growth as the markets evolve