High Frequency Trading: What is it Good For?

Speed and Latency in U.S.
Equity Markets
Austin Gerig
Division of Economic and Risk Analysis (DERA)
US Securities and Exchange Commission
The Securities and Exchange Commission, as a matter of policy,
disclaims responsibility for any private publication or statement by any of
its employees. The views expressed herein are those of the author and
do not necessarily reflect the views of the Commission or of the author’ s
colleagues upon the staff of the Commission.
Overview
• Part I
– Regulation of Clock Sync in U.S. Equity
Markets
• Part II
– Speed and Latency in U.S. Equity Markets
Why Important/Interesting?
• Clock Sync
– Allows sequencing of geographically separated market events
• Speed/Latency
– Better queue position in order books
– First response to news events
– Fast arbitrage
• Portfolio arbitrage (Futures to equities, ETF’s, etc.)
• Latency arbitrage (Flash Boys)
• Latency Fluctuations
– Determination of non-compliance with regulation
– Determination of when data feeds are unreliable
Part I: Regulation of Clock Sync
FINRA Rule for Broker/Dealers
Firms have until Feb 17, 2017 or Feb 20, 2018 to apply the new clock sync
standard (the current standard is within 1 second of NIST).
Sources: FINRA manual, Rule 4590 and http://www.finra.org/industry/faq-oatsclock-synchronization-faq
SIP Plan Requirements for
Exchanges
• “Exchanges use a clock sync methodology
ensuring that timestamps are accurate
within tolerances of 100 microseconds or
less.”
• source:
http://www.nasdaqtrader.com/content/new
salerts/2014/utp/utp2014-06.pdf
Upcoming CAT Requirements
Implemented in phases between 2017-2019.
Source: Approved CAT NMS Plan http://www.catnmsplan.com/
Part II: Speed and Latency in U.S.
Equity Markets
Different types of latency
• Latency of algorithmic strategies
– FPGA (~1us)
• Latency from trading system to matching
engine
– Co-location (~25-50us)
• Latency between trading venues*
– RF - Laser/Microwaves (~100-8000us)
*Focus of the rest of the talk
Preliminaries
• Exchange-to-SIP Latency
– The time elapsed between a message being sent from an exchange,
and it being received by a Securities Information Processor (SIP)
– Currently recorded in microseconds (a blink of an eye takes 150
milliseconds or 150,000 microseconds)
• SIP
– CTA (Tape A and Tape B securities) and UTP (Tape C securities)
• TAQ data (trade and quote data)
– Publicly available data containing SIP messages: equity trades and
quotes.
– Each trade/quote entry contains information regarding when the
message was sent by an exchange to a SIP, and when the message
was received by a SIP
Minimum Exchange-to-SIP Latencies
Figures: Speed of light latency from NYSE/ARCA , BATS/Direct Edge, and IEX to and
from NASDAQ. Left: Direct line of sight through air. Right: Following roads through
fibre optic cable. Source: Internal calculations and Google maps.
Measured vs. Minimum Latency
Source: TAQ data and calculations from previous slide.
Latency is not constant, it
fluctuates intraday and over time
NASDAQ to UTP SIP latencies for AAPL Trades. Source: TAQ data.
Pareto type IV distribution best fits
intraday fluctuations
Distribution of NASDAQ to UTP SIP latencies for AAPL Trades. Source: TAQ data.
Exchange to SIP latency distribution
Pareto IV parameters
Best fit parameters for Pareto IV distribution of exchange to UTP SIP latencies for
AAPL trades. Source: TAQ data.
Exchange to SIP latency distribution
Pareto IV plots
Source: TAQ data.
UTP SIP upgrade on 10/24/16
Source: https://globenewswire.com/newsrelease/2016/10/24/882097/0/en/Securities-Information-Processor-SIP-Migrates-tothe-Nasdaq-Financial-Framework-and-INET-Technology.html
Observed changes for BATS to
UTP SIP latency due to upgrade.
Source: TAQ data.
Clock drift is observable
Source: TAQ data.
Conclusions
• Latencies between trading venues fluctuate intraday and through
time (they are not a single value)
• Pareto IV distribution best fits these fluctuations
• Exchanges have different parameters for the distribution, but those
parameters are interpretable and related to:
– geographic distances
– technological infrastructure
– possible universal behavior of network delay
• Technological upgrades can have significant effects on latency
• Clock drift contributes to latency fluctuations and is observable