Quantifying the High-Frequency Trading “Arms Race”

In the past few decades, financial markets across most major asset classes—equities, futures, treasuries, currencies, options, etc.—have transformed from human beings interacting with each other on trading floors, pits and desks, to computerized trading algorithms interacting with each other in exchange computer servers. On the whole, this transformation from humans to computers has brought clear, measurable improvements to various measures of the cost of trading and liquidity, much as information technology has brought efficiencies to many other sectors of the economy. But this transformation has also brought considerable controversy, particularly around the importance of speed in modern electronic markets.

This study uses a simple new kind of data and a simple new methodology to study the phenomenon at the center of the controversy over speed: latency arbitrage. In words, a latency arbitrage is an arbitrage opportunity that is sufficiently mechanical and obvious that capturing it is primarily a contest in speed. Conceptually, a latency arbitrage is an economic rent from symmetric public information—information that, in principle, is disseminated to the whole market simultaneously and publicly, so should not be a source of arbitrage profits.

The data are the “message data” from an exchange, as distinct from widely familiar limit order book datasets. Limit order book data provide the complete play-by-play of one or multiple exchanges’ limit order books, often with ultra-precise timestamps. But what is missing are the messages that do not affect the state of the order book, because they fail. For example, if a market participant seeks to snipe a stale quote but fails—their immediate or cancel (IOC) order is unable to execute immediately so it is instead just canceled—their message never affects the state of the limit order book. Or, if a market participant seeks to cancel their order, but fails—they are “too late to cancel”—then their message never affects the state of the limit order book. But in both cases, there is an electronic record of the participant’s attempt to snipe, or attempt to cancel. And, in both cases, there is an electronic record of the exchange’s response to the failed message, notifying the participant that they were too late.

Our method relies on the simple insight that these failure messages are a direct empirical signature of speed-sensitive trading. If multiple participants are engaged in a speed race to snipe or cancel stale quotes, then, essentially by definition, some will succeed and some will fail. The essence of a race is that there are winners and losers—but conventional limit order book data doesn’t have any record of the losers.

We obtained from the London Stock Exchange all message activity for all stocks in the FTSE 350 index for a 9 week period in Fall 2015. Using this data, we can directly measure the quantity of races, provide statistics on how long races take, how many participants there are, the diversity and concentration of winners/losers, etc. And, by comparing the price in the race to the prevailing market price a short time later, we can measure the economic stakes, i.e., how much was it worth to win.

The main results of our study are as follows:

• Races are frequent. The average FTSE 100 symbol has 537 latency-arbitrage races per day. That is about one race per minute per symbol.
• Races are fast. In the modal race, the winner beats the first loser by just 5-10 microseconds, or 0.000005 to 0.000010 seconds. In fact, due to small amounts of randomness in the exchange’s computer systems, about 4% of the time the winner’s message actually arrives to the exchange slightly later than the first loser’s message, but nevertheless gets processed first.
• A remarkably large proportion of daily trading volume is in races. For the FTSE 100 index, about 22% of trading volume and 21% of trades are in races. Cochrane (2016) describes that trading volume is “The Great Unsolved Problem of Financial Economics.” Our results suggest that latency arbitrage is a meaningful piece of the puzzle.
• Races are worth just small amounts each. The average race is worth a bit more than half a tick, which on average comes to about 2GBP.
• Race participation is concentrated. The top firms disproportionately snipe. The top 3 firms win about 55% of races, and also lose about 66% of races. For the top 6 firms, the figures are 82% and 87%. In addition to documenting concentration, we also find that the top 6 firms are disproportionately aggressive in races, taking about 80% of liquidity in races while providing about 42% of the liquidity that gets taken in races. Market participants outside the top 6 firms take about 20% of liquidity in races while providing about 58%. Thus, on net, much race activity consists of firms in the top 6 taking liquidity from market participants outside of the top 6.
• In aggregate, these small races add up to a significant proportion of price impact and the effective spread, key microstructure measures of the cost of liquidity. Price impact from trading in races is about 31% of all price impact, and about 33% of the effective spread. This suggests latency arbitrage deserves a place alongside traditional adverse selection as one of the primary components of the cost of liquidity.
• Market designs that eliminate latency arbitrage could meaningfully reduce the market’s cost of liquidity. We find that the latency arbitrage tax, defined as the ratio of daily race profits to daily trading volume, is 0.42 basis points if using total trading volume, and 0.53 basis points if using only trading volume that takes place outside of races. The average value-weighted effective spread paid in our data is just over 3 basis points. We show formally that the ratio of the non-race latency arbitrage tax to the effective spread is the implied reduction in the market’s cost of liquidity if latency arbitrage were eliminated; that is, if liquidity providers did not have to bear the adverse selection costs associated with being sniped. This implies that market designs that eliminate latency arbitrage, such as frequent batch auctions, would reduce investors’ cost of liquidity by 17%.
• These small races add up to a meaningful total “size of the prize” in the arms race. The relationship between daily latency-arbitrage profits and daily volume is robust, and indeed the latency-arbitrage tax on trading volume is roughly constant in our data. Using regression analysis, we find that the annual sums at stake in latency arbitrage races in the UK are about GBP 60 million. Extrapolating globally, our estimates suggest that the annual sums at stake in latency-arbitrage races across global equity markets are on the order of $5 billion per year.

Whether the numbers in our study seem big or small may depend on the vantage point from which they are viewed. As is often the case in regulatory settings, the detriment per transaction is quite small: the average race is for just half a tick, and a roughly 0.5 basis point tax on trading volume certainly does not sound alarming. But, because of the large volumes, these small races and this seemingly small tax on trading add up to significant sums. A 17% reduction in the cost of liquidity is undeniably meaningful for large investors, and $5 billion per year is, as they say, real money—especially taking into account the fact that our results only include equities, and not other asset classes that trade on electronic limit order books such as futures, treasuries, currencies, options, etc. Overall, the latency arbitrage tax does seem small enough that ordinary households need not worry about it in the context of their retirement and savings decisions. Yet at the same time, flawed market design drives a significant fraction of daily trading volume, significantly increases the trading costs of large investors, and generates billions of dollars a year in profits for a small number of HFT firms and other parties in the speed race, who then have significant incentive to preserve the status quo.

The complete paper is available for download here.

The authors have also provided a detailed code package for other researchers, regulators or practitioners to conduct their own studies using message data, available here.