The NYSE carefully reviewed the recently published report from MITRE entitled, " Scaling of Inefficiencies in the U.S. Equity Markets: Evidence From Three Market Indices and More Than 2,900 Securities." While we appreciate the complexity of the data and calculations required to produce the report, we believe there are serious flaws in the methodology. These issues lead the report to overstate the scale of the alleged issue.
The report investigates "latency arbitrage opportunities and realized opportunity costs" from trading in stocks in the Russell 3000® index during 2016. The paper attempts to identify trades that occur when different data feeds show different prices at the same time. The paper estimates that 23.71% of all trades occur when the National Best Bid and Offer (NBBO) calculated from exchange proprietary data feeds is different from the NBBO published by the Security Information Processors (SIPs). The paper estimates that these trades result in an annual "realized opportunity cost (ROC)" of greater than $2 billion. This result is substantially different than the Bartlett and McCrary paper (2017)1, which estimated that the SIP and direct feeds showed a different price for just 3% of all trades and that potential opportunity costs were an order of magnitude lower.
We believe there are a few key issues leading to an inaccurately high cost estimate.
Most importantly, the authors based their ROC calculations on the quantity of each trade without consideration of the actual quoted quantities available on the data feeds. Consider the following example:
The authors informed us that they calculate the ROC on the 200 shares executed at the SIP NBB of $10.00, or $0.01 per share, resulting in a $2.00 ROC. The paper assumes both quotes are equally accessible at the same instant, but disregards the fact that 200 shares could not have executed at $10.01 on Nasdaq; Nasdaq was only showing 100 shares available on its direct feed at the better price.
Extending this example, if multiple trades executed at $10.00 on different venues, before the SIP quote was updated, each of these would be counted towards the ROC. This implies that potentially any number of shares were available for arbitrage, when there were really only 100 shares that could offset any trade. For there to be an arbitrage opportunity, one must be able to execute the opposite leg of the full quantity of the trade.
Our second key issue is that we are not clear how the authors handled odd lots quoted on direct feeds. If they calculated an NBBO based on odd lots they would further exacerbate the over-estimation described above. The authors could avoid this problem by looking for executions in odd lot size at the prices quoted on the direct feeds. If the odd lot trades occur, it is clear that the shares are no longer available and any subsequent trades at the SIP NBBO could not possibly be afforded an arbitrage opportunity.
Third, we learned from the authors that they sourced the unified SIP and proprietary feed data from Thesys based on data they collected in Carteret, NJ. We believe running such a study from a more neutral geographic location, such as Secaucus, NJ, would lead to significantly fewer perceived arbitrage opportunities. While the authors took a measured approach to try an account for the geographic latencies between the three New Jersey exchange data centers, those latencies are not static and can vary based on a variety of factors, such as spikes in market activity and telecommunications performance.
Fourth, the executions used to calculate potential ROC were sourced from the SIP rather than from the proprietary feeds. Exchange proprietary feeds publish executions in addition to quotes, and we believe that using executions from these proprietary feeds would indicate far fewer potential arbitrage opportunities.
Due to these issues, NYSE concludes that the results presented in the paper are flawed and do not reflect actual latency arbitrage opportunities and costs.
1. How Rigged Are Stock Markets? Evidence from Microsecond Timestamps, Robert P. Bartlett II, Univ. of California, Berkeley and Justin McCrary, Univ. of California Berkeley, Columbia University, NBER