How Rigged Are Stock Markets? Evidence From Microsecond Timestamps
Robert P. Bartlett III
University of California, Berkeley – School of Law; University of California, Berkeley – Berkeley Center for Law, Business and the Economy
University of California, Berkeley; National Bureau of Economic Research (NBER)
Joel Greenblatt Owned Hedge Fund On Why Value Investing Isn’t Working Now
Acacia Capital was up 12.27% for the second quarter, although it remains in the red for the year because of how difficult the first quarter was. The fund is down 14.25% for the first half of the year. Q2 2020 hedge fund letters, conferences and more Top five holdings Acacia's top five holdings accounted for Read More
July 20, 2016
We use new timestamp data from the two Securities Information Processors (SIPs) to examine SIP reporting latencies for quote and trade reports. Reporting latencies average 1.13 milliseconds for quotes and 22.84 milliseconds for trades. Despite these latencies, liquidity-taking orders gain on average $0.0003 per share when priced at the SIP-reported national best bid or offer (NBBO) rather than the NBBO calculated using exchanges’ direct data feeds. Trading surrounding SIP-priced trades shows little evidence that fast traders initiate these liquidity-taking orders to pick-off stale quotes. These findings contradict claims that fast traders systematically exploit traders who transact at the SIP NBBO.
How Rigged Are Stock Markets? Evidence From Microsecond Timestamps – Introduction
A pressing question for contemporary equity markets concerns the extent to which they systematically favor fast traders at the expense of slower ones. As vividly described in Michael Lewis’ Flash Boys, these concerns center on the differing speeds with which traders can access data emanating from one dozen U.S. stock exchanges and approximately fifty non-exchange trading venues. Because trading rules benchmark trade execution to the national best bid or offer (the “NBBO”) available across exchanges, traders with faster access to a venues’ quotes can effectively foresee the NBBO on which slower traders rely. Indeed, market rules conventionally encouraged brokers to fill market orders at (or better than) the NBBO as it appears on the two centralized Securities Information Processors (“SIPs”) to which all exchanges are required to report updates to their best bids and offers. Yet the latency with which the SIPs process the NBBO relative to a fast trader potentially allows fast traders to anticipate price changes in the SIP-generated NBBO. If retail and other investors look to trade at the SIP NBBO, this speed advantage accordingly creates risk-free opportunities as fast traders can choose to trade or not trade at NBBO prices they know to be stale.
Until recently, understanding the extent to which these opportunities actually arise within the market has been hampered by the absence of data concerning the latency of the SIPs relative to quotation data obtained directly from exchanges through their proprietary data feeds. In this paper, we use new timestamp data provided by the two SIPs to conduct the first systematic analysis of the latency with which the SIPs process quote and trade data. This new timestamp data, which exchanges and broker-dealers have been required to report to the appropriate SIP in their mandatory transaction reports since August 2015, provides the precise time (measured in microseconds) at which a trading venue either updated a quotation or executed a trade.1 Moreover, amendments to the SIP operating procedures at this time obligated the two SIPs to record in microseconds the precise time at which each SIP processed a trade or quotation update submitted by an exchange or broker-dealer.2 Comparing these two timestamps thus permits a direct analysis of the SIP processing latency for all trades and quote updates across the entire market.
We take advantage of these newly released data with microsecond timestamps to estimate quote and trade latencies for all stocks within the Dow Jones 30 during the first nine months of these new reporting requirements. Overall, we show that the SIP latency in processing both quote updates and trades is, on average, virtually non-existent for any human trader. This is especially true with respect to quote updates. We find that the mean time gap between the time a quote update is recorded by an exchange matching-engine and the time it is processed by a SIP is a scant 1.13 milliseconds. Mean latency for processing trades, however, is approximately 20 times higher, clocking in at 22.84 milliseconds. The slower processing time for trades largely reflects the fact that nearly one-third of trades occur in non-exchange venues whereas quote updates are disseminated by exchange matching engines. Excluding trades executed in nonexchange venues, mean reporting latency for trades is less than 1 millisecond. While latencies are small on average, we document long right-hand tails for both quote and trade reports. For instance, more than 2% of all quote updates in our sample transactions from the Nasdaq BSX and the Chicago Stock Exchange had latencies exceeding 10 milliseconds.
Confirming the quality of these new measurements, we document that the variation exhibited by quote and trade latencies reflects the institutional structure of SIP-reporting obligations. For example, transaction reports in securities listed on the NYSE (Tape A securities) and in securities listed on any regional exchange (Tape B securities) must be sent to the NYSE-SIP located in Mahwah, New Jersey, while those in securities listed on Nasdaq (Tape C securities) must be sent to the Nasdaq-SIP located 35 miles south in Carteret, New Jersey. For Tape A securities, this requirement means quote and trade reports released by the NYSE matching engine arrive at the NYSE-SIP almost instantaneously, while quote and trade reports in Tape A securities occurring on the Nasdaq matching engine are certain to arrive at the NYSE-SIP at least 188 microseconds after they occur (the approximate time it takes light to travel 35 miles). The data confirm this basic institutional prediction, strengthening confidence in the quality of measurement.
In addition to documenting basic descriptive information about the SIPs’ reporting latencies, we utilize the new timestamps to explore empirically the economic significance of these latencies for liquidity takers such as retail traders. We focus on two aspects of SIP reporting latency that are often cited as potential problems and that are presently at the center of Department of Justice investigations into retail market-making firms Citadel and KCG (Levinson, 2016). The first of these conventional wisdoms holds that brokers filling marketable orders at (or within) the SIP-generated NBBO may do so at stale prices to the disadvantage of retail investors. For instance, suppose a direct feed showed the NBBO changing from $10.01 x $10.00 to $10.00 x $9.99, while the SIP’s NBBO remained at $10.01 x $10.00. A broker might fill buy orders by selling to them at $10.01 (the stale NBO reflected in the SIP NBBO) rather than at $10.00 (the NBO shown in its direct feed). The second of these conventional wisdoms holds that in publishing trade execution statistics, market centers might choose as their pricing benchmark the slower SIP-generated NBBO to enhance artificially their performance metrics. Returning to the example above, by using the SIP NBO of $10.01, the broker would report an effective spread of just $0.01 (twice the difference between the trade price of $10.01 and the midpoint of the SIP NBBO) rather than the actual effective spread of $0.03.
See full PDF below.