Does Your Algorithm Contain A Ticking Time(r) Bomb?
I hesitated using the word “tick” in the title of this post, lest potential readers think I am writing yet another post on tick sizes. But I assure you, this post has absolutely nothing to do with tick size.
Rather, this post covers a topic that is rarely discussed, but can have an outsized impact on performance, namely how the use of timers in sell-side algorithms can wreak havoc on passive trading performance.
Before jumping in, let's first discuss the results of a recent academic paper on high frequency trading, since this will help put our discussion of algo timers in context.
Trading has become fast. Wicked fast.
A recent study by researchers from the UK FCA and the University of Chicago estimates that the “latency arbitrage tax” imposed by high frequency traders is approximately 0.42 bps. While small on a per trade basis, these costs are actually large when aggregated across trades. In fact, the others estimate that the total latency arbitrage tax paid across global equity markets could be upwards of 5 billion dollars per year.
The study contains interesting insights into the “latency arbitrage races” that occur periodically throughout the day, where HFTs race to pick off stale quotes. The authors find that these races happen about once per minute, with multiple firms participating in each race. However, only a small number of traders generally “win”. In fact, 80% of the races were won by the same six firms in their study – and the winners beat the “fastest loser” by only 5-10 microseconds, on average. This difference may explain why a prominent HFT firm complained when the NYSE proposed building a wireless tower that would reduce latency by a whopping 2 microseconds.
While we all wait to see if the SEC will do anything to reduce this “tax” – either directly via changes in regulations or indirectly by allowing exchanges to impose countermeasures, like speedbumps – buyside traders are caught in the crossfire. Given the substantial costs of entering the low-latency arena, most buyside traders rely on sell-side tools, such as smart routers and algorithms to compete for liquidity. These broker tools often utilize the same tools employed by HFTs, such as direct market feeds, co-location, wireless transmission, etc. But even then, the sell-side is at a disadvantage.
Some of this disadvantage is structural. Most sell-side firms, for example, co-locate at a single venue – typically NY4 in Secaucus. While this is helpful when routing to co-located exchanges like BATS-Y, EdgX, etc., sell-side brokers incur latency when routing to other venues, like the Nasdaq and NYSE. Of course, sell-side firms could co-locate at multiple venues. But this often is not useful, since agency algorithms typically have constraints that HFTs don’t have. For example, an algorithm working a 100-share quantity cannot post that 100-shares at multiple venues, even if only fleetingly. Consequently, for an algorithm to move quantity from Exchange A to Exchange B, it must first get an “out” from Exchange A before routing to Exchange B. In aggregate, then, information must effectively flow from Exchange A to Exchange B, and that incurs latency that simply cannot be avoided when “smart routing” across multiple venues.
The Ticking Time(r) Bomb in Algorithms
With that said, not all incremental latency incurred by sell side algorithms is unavoidable. One tremendous avoidable source of latency stems from algo provider’s decision to use “timer-based’ instead of event-based’ processing. Here, timer-based is not meant to refer to the timers that keep TWAP and VWAP (and potentially other algorithms) on schedule. Use of timers in that context is natural and generally benign, as they drive the pacing of child orders into the market. Rather, we are talking about how these child orders themselves are managed after they have been routed into the market.
For example, after the VWAP algorithm has decided to buy 100 shares passively, the algorithm determines a price and venue for that child order, and then routes it out. From that point on, the algorithm must monitor that order and update its price and/or venue as necessary. A timer-based algorithm “wakes up” periodically, e.g., every 500 milliseconds, samples the state of the market, and then takes action before beginning another 500-millisecond slumber. All events that occur while the algorithm snoozes are essentially put on hold (or ignored entirely) until the algorithm wakes again from its slumber.
Consequently, timer-based algorithms can lead to subpar performance both because of its increased adverse selection risk as well as its inefficient queue management. For example, when the timer-based algorithm eventually wake-up after its siesta, it may learn that its limit orders have already been “picked off” as its stale limit price became fodder for HFTs. Or the algorithm may learn that prices have since moved away, in which case the algorithm updates its price to the new price level, likely joining an already lengthy queue at the new price.
Three Things You Can Do
Ask your broker
All else equal, buyside traders should seek out algorithmic providers who use event-based processing. Event-based algorithms are always on, listening for quote changes, trades, the order’s own executions and cancelation, etc. in real-time so they can respond immediately, when necessary. In event-based systems, speed is of the essence. For them to be effective, they must receive market data, fills, cancels, etc. as quickly as possible and then implement those decisions quickly. For such systems, investments in low latency market data, co-location, wireless connectivity, etc. can have a meaningful impact. Of course, this doesn’t mean that sell-side firms would be (or should be) on par with the fastest HFTs. But it will definitely give some algorithmic providers a distinct advantage over other algorithmic providers.
Therefore, a starting point is simply to ask your brokers whether they are event-based or timer-based when it comes to processing child order information. This question can provide useful insights into the algorithm’s potential for reducing adverse selection, queue management, etc. Indeed, this should be a standard question on the “broker questionnaires” that many buyside firms require their brokers to complete.
Of course, broker performance should still be evaluated empirically and on an ongoing basis, since having event-based processing alone doesn’t ensure good performance. Making bad decisions fast is not necessarily better than making good decisions more slowly! But a timer-based system almost surely cannot compete with the best-of-breed event-based algorithms.
Focus on relevant algo enhancements
Furthermore, knowing whether an algorithm is event-based or timer-based can help put a broker’s performance in context and help identify ways to enhance performance – and discount changes that hinge on fast reaction times. For example, poor performance driven by adverse selection cannot be easily addressed in a timer-based system. Simply speeding up other parts of the system will likely have a negligible effect. Even enhancing the underlying pricing and routing models may have limited effect if the performance of these enhanced models hinge on timely market data, as they likely do. But your broker may be able to take actions that can help correct the consequences of timer-based algorithm. For example, increasing the use of inverted venues for resting, for example, may be helpful, as they allow the algorithm to jump those queues that may have already formed while the algo was peacefully napping. This way, the algorithm is more likely to fill before prices have had a chance to move.
Again, while event-based processing is better able to leverage a low latency infrastructure than timer-based processing, we’re not saying that low latency investments are entirely useless in a timer-based system. On the margin, it is better to cut latency wherever possible. But it is sort of like how I bought an expensive Les Paul guitar a few years back. It’s helpful in that it shows my friends that I’m serious about “my music”, but it doesn’t really help make my rendition of “Smoke on the Water” sound that much better.
The author is the Founder and President of The Bacidore Group, LLC. For more information on how the Bacidore Group can help improve trading performance as well as measure that performance, please feel free to contact us at email@example.com or via our webpage www.bacidore.com.
For an overview of our other services, please click HERE.
And please check out our other blog posts available HERE.
And if you'd like to receive notification of new posts, please join our mailing list below.
Copyright 2020, The Bacidore Group, LLC. All Rights Reserved.
 If you missed them, the initial post and its follow up can be found here and here.
 See Aquilina, Budish, and O’Neill (2020), which can be found here. A summary of the paper can be found here.
 See Alex Osipovich’s article “NYSE Antennas Spark High-Speed Trader Backlash”, Wall Street Journal, Aug. 8, 2019.
 For a discussion of the proposed EdgA speedbump and the innovative IEX D-limit order, see our posts here and here, respectively.
 Of course, if the algorithm is parsing orders out in a highly deterministic pattern, e.g., at fixed intervals and/or precisely on “round” milliseconds, timers could potentially lead to gaming.