Not So Fast Nanex. Debunking the HFT Ford Study
There has been a great deal of discussion recently about whether financial markets are rigged in favor of dark pools, high frequency traders and the large banks. I’ve been following this topic and the material emerging on all sides of the debate, and one of the most interesting aspects is related to liquidity.
I am planning to do a series of blogs exploring various aspects of liquidity. These blogs will include reviews of academic papers and research on the topic, a look at what other markets are doing and also discussions of current events. Please comment if you have additional insights or analysis that you have performed on this topic. I look forward to engaging with you on the topic of liquidity and how it influences and works within the market.
Recently a gentleman by the name of Eric Hunsader (he’s founder of market data company Nanex) published a study talking about what took place during 4.6 milliseconds on July 11, 2014 when a trader attempted to buy 20,000 shares of Ford (F). Hunsader concluded that the incomplete trade was evidence of a rigged market and has promoted his study broadly on social media. He even got the attention of Michael Lewis, author of Flash Boys and self‐appointed monarch of the “market is rigged” camp.
Not so fast. I decided to take this study and attempt to re‐create what happened using industry standard, publicly available information. This blog includes my analysis of what happened before, during, and after those 4.6ms. It seems that Hunsader and Nanex’s data tools aren’t quite up to the task of seeing what really happened. According to my analysis, the market is far from rigged.
The buyside trader (Hunsader has not identified the trader, but only said that s/he came to Nanex to study what happened) attempted to execute a trade to buy 20,000 shares of Ford. The algorithm/router performing this trade took over 4.6ms, but successfully obtained over 12,000 shares at the desired price. During this time 12 trades occurred across seven exchanges and there are two notable gaps in trading comprising almost 2ms at the beginning of the series.
Nanex concluded that High Frequency Trading (HFT) is the reason the order was only partially filled—and therefore “the stock market is rigged.” His reason: the “order cancellations (that) happen(ed) far faster than trade executions …. before and during the trader’s order (and) were not a coincidence.” Amping up the rhetoric even more, Hunsader called what happened “premeditated, programmed theft.”
I reviewed Nanex’s analysis of the market in Ford during the time period in question using industry standard, microsecond resolution timestamps from publicly available, direct market data feeds to see what really happened. It was immediately clear that there were some fatal flaws in Nanex’s analysis. One obvious cause for these errors is that Nanex’s analysis did not use the same direct market data feeds used by the Buysider’s router. Ironically, by insisting on only using the SIP (instead of industry standard, direct feeds) in all of its analysis, Nanex’s views of micro‐market structure events are generally flawed.
Nanex’s study included the below chart on the left marked Exhibit 1a (“Nanex”). To provide additional detail, I’ve enlarged the diagram in Exhibit 1a by increasing the time resolution and adding real microsecond timestamps in Exhibit 1b (“Truth”). I have also added the light blue vertical lines, which mark the 12 executions reported during the 4.6 millisecond episode (between 09:47:56.5694 and 09:47:56.574). This will allow us to take a closer look at the correlations between the various charts and trade times.
Nanex asserts that “HFT reacted faster than the original order could be routed to other exchanges and beat our trader to those shares.” This claim is clearly incorrect. As the chart above reflects, there was a 0.95 millisecond gap in time after the first order was executed where you see no trades and no offers cancelled. The large time gaps (yellow boxes) between the 1st/2nd and 2nd/3rd executions suggest that poor routing decisions impacted execution quality — not an HFT or a liquidity problem. So why did this happen?
Fee Sensitive Routing: In Exhibit 2 above, you can see that the routing strategy deployed by the Buysider acted in a sequential manner that was likely incentivized by the fee/rebate applicable at each venue—rather than employing a “smarter” router that was optimized for execution quality (i.e. removing the available resting offers). In the actual trading sequence, the routing strategy deployed first removed all displayed liquidity on BX (Boston), where the executing broker would have received a rebate of 4 to 15 mills per share. The Buysider’s router then waited almost a full millisecond before removing all displayed liquidity on EdgeA, where the executing broker would have received a rebate of 2 mills per share. It seems reasonable to assume that the Buysider’s router hastened the price level transition when it swept the entire inside offers on BX and EdgeA. It also seems reasonable to assume that the Buysider’s order would have been completely filled had it simultaneously routed to “pay to remove” venues (where there was sizeable resting offers) when it swept the insider offer on EdgeA. The end result of a “smarter,” less fee sensitive router would have been the removal of all of the resting insider offers on BATS, EdgeX, Nasdaq, ARCA and NYSE instead of the partial fills received.
From the timestamps, it appears that the orders from the Buysider for BATS, EdgeX and NYSE (Trades #3, #4, #5) all hit the exchanges at nearly the same time (after adjusting for approximate network latency), which suggests the router had the sophistication to hit all exchanges at the same time if they really wanted to. Further, Nanex mistakenly claims that “none of this would be possible if the direct feeds weren’t illegally supplying HFT with faster information than the SIP;” however, the SIP updated the EdgeA quote for Trade 2 at 09:47:56.571300, which would have given any trader (regardless of whether they were using “illegal” direct market feeds or the SIP) enough time to cancel offers at BATS, EdgeX, and NYSE (Trades #3, #4, #5) in response to the market impact — if that was the intent.
Exhibit 3 is from Nanex’s “update” to its “research” and is the result of an order executed using a more efficient router, which was supplied by the IEX ATS. In this example, the Buysider reported better execution quality. This further supports the facts that the incomplete execution experienced by the Buysider (Exhibit 1 and Exhibit 2) is simply the result of the Buysider using a less-than-smart order router—not the boogie man or “HFT” or a “rigged” market. Indeed, I would be curious to hear Nanex’s rationalization of how trades like Exhibit 3 can exist in a “rigged” market.
Before the Trade
Before the Buysider’s order enters the market, the relative sizes of the $17.37 bid and the $17.38 offer reflect a large imbalance. In Exhibit 4 below, you can see that the $17.37 bid size is almost double the $17.38 offer size and that the offer size almost doubles in size by the end of the 4.6ms period. The growth of the $17.37 bid size, in conjunction with the fading $17.38 offer size, illustrates a normal price level transition that occurs when the demand exceeds the supply — not “phantom liquidity” or “irrefutable evidence” of HFT front-running. Again, it is important to understand that this transition was in motion before the Buysider’s first trade was executed.
After the Trade
Nanex claims that front-running allowed HFT to artificially raise the price of the underlying security and eroded price stability. The facts, illustrated in the chart below in Exhibit 5 (which depicts all Ford (F) trading activity for the subsequent 30 seconds after the Buysider’s purchase described above) shows that all but 200 shares over this time frame traded at the exact price the purchaser had sought to pay ($17.38).
In sum, contrary to claims by market expert Michael Lewis that Hunsader offers “clean, simple, irrefutable” proof that the markets are rigged, the data actually supports the opposite conclusion. It should also call into question the efficiency of the router being used by the Buysider and a general misunderstanding of the how the market works.
This example and analysis demonstrated the efficiency of a properly functioning market: There was a market imbalance (i.e., there was more interest on one side of the book than the other — in fact almost double as much interest on the bid) and when the Buysider began removing offers in an inefficient way, the offered price eventually did go higher. This example and analysis also demonstrates the importance of using the right order router: A more efficient router likely would have removed more (or all) of the displayed liquidity — as demonstrated in Exhibit 3 — even if that meant prioritizing “pay to remove” venues ahead of “rebate to remove” venues in its liquidity seeking decisions .
Ultimately, Nanex’s study merely tells us what most traders already know—that the market reacts to new information by changing price, and if you don’t choose the right “smart” router, your execution quality will suffer.