Market Mechanism Refinement on a Continuous Limit Order Book Venue HAYDEN MELTON Deakin University & Thomson Reuters (Markets) LLC

This letter describes an exercise in market mechanism refinement that was recently undertaken on a major electronic trading venue: Thomson Reuters Matching. The goal of the exercise was to establish a minimum response time for market participants on the venue while simultaneously seeking to avoid other unrelated impacts on their trading strategies. The design of the refinement is described, and some of its consequences are discussed. Categories and Subject Descriptors: K.4.4 [Computers and Society]: Electronic Commerce

1.

INTRODUCTION

Thomson Reuters Matching (TRM) is one of two major interbank electronic trading venues for spot foreign exchange (FX) [Hasbrouck and Levich 2017]. It has been in wide-use since its launch in 1992—oftentimes trading in excess of $100 billion USD daily of spot FX—and up until mid-2016 its market mechanism was a continuous limit order book with price-time priority matching. Problems with this general mechanism have recently been identified and are described in detail elsewhere, but to summarize, its processing of messages in the temporal order in which they are received: (1) causes an ongoing, socially wasteful technology ‘arms race’ among participants where each seeks to be marginally faster at responding than their peers [Harris 2013; Budish et al. 2015], and (2) exacerbates the extent to which seemingly inescapable informational asymmetries in market data distribution from and message processing times by a venue determine winners and losers when participants race for prices on it [Melton 2017]. To address these two problems—the first being one of social welfare and the second being one of fairness—in mid-2016 the mechanism by which TRM matches orders was altered so as to deemphasize the effect of speed on the venue. Architecturally, the way in which this was accomplished is illustrated in Fig.1. In particular, a new module, the Ideal Latency Floor [Melton 2015], was introduced into the matching engine component of the venue. At a high-level the way this new module works is by ‘intercepting’ messages before they reach the existing pricetime matching algorithm, and reordering them. The specific procedure by which this module reorders them is described in Section 2, but its overall effect is that at short timeframes messages are generally not processed for matching in the temporal order in which they are received, while at longer timeframes messages received Author’s address: [email protected] & [email protected] Views expressed herein do not constitute legal or investment advice, and do not necessarily reflect those of the author’s employer. Under review, May 2017

2

·

H. Melton

Fig. 1.

Architecture of an Electronic Trading Venue and its Refinement.

earlier continue to be processed for matching before those received later. In this way the module can be thought of as providing only a slight departure from the existing price-time matching algorithm implemented by the venue. The term refinement is thus used quite deliberately here to convey the intended subtlety of this departure. Radical departures from long-established market mechanisms may be undesirable due to the risks they pose to both participants and venue operators [Phelps et al. 2010]. In particular, such departures may render participants’ existing trading strategies—in which they have likely made considerable investments over long periods of time—unprofitable, thereby causing them to cease their trading activities on the venue, which is not an outcome venue operators desire. As is a focus of Section 3, it is therefore prudent to scrutinize the effects a proposed change to a mechanism may have in terms of its impacts on participants’ existing trading strategies, with a view to minimizing impacts that do not relate to the goal of the change. 2.

DESIGN OF THE REFINEMENT

The specific goal of the refinement to TRM can be stated more precisely than “deemphasizing the effect of speed”. It was actually to establish a minimum response time such that all participants who could respond to an event within that time would have an equal probability of winning a race to make or take a price on the venue, when they so choose to compete in one. Put another way, responsive to a particular stimulus (e.g., a specific market data update), all participants who can send an order message to the venue within T milliseconds of observing that stimulus should have an equal chance of winning the race for the resource on the venue for which their order message is competing. If the events that trigger races among participants for resources on a venue could be known by the venue operator, a reasonable approach to establishing a minimum response time T on the venue for those events might be as follows: (1) upon the observation of such an event, start a timer that shall run for T milliseconds; (2) before the timer has reached T , collect into a buffer all the orders received by the venue responsive to that event, so as to ‘temporally aggregate’ those orders; (3) upon the timer reaching T , shuffle the orders in the buffer before draining them from it for matching. The problem with the above approach is that the events that cause participants to race for resources on a venue are, for a number of reasons, not known to the venue operator. In particular, the FX market is fragmented and the operator of one venue generally does not have access to events (e.g., market data updates) emanating from other, competing venues. Further, participants trading strategies are among their Under review, May 2017

Market Mechanism Refinement on a Continuous Limit Order Book Venue

·

3

most closely-guarded trade secrets—even if the superset of all events that might cause orders to be sent could be known to a venue operator, the specific event triggering the sending of an order could not be known by the operator with any certainty. For the above reasons a degree of satisficing [Simon 1996, ch.2] had to occur in the design of the refinement. What was selected as the proxy for the events that cause the sending of order messages were those order messages themselves. Obviously the order messages sent to a venue are always observable by the operator of that venue. Similarly, and also somewhat obviously, races for prices on a venue must also necessarily manifest in the form of order messages being received by that venue. Order messages, it seems then, are a good proxy for the events that trigger them. The specific way in which the event-based approach described earlier was adapted for its implementation on TRM using order messages as a proxy for events is as follows. Upon receipt of an order message (cf. event in the earlier approach), its limit price is compared to either the prevailing best bid or offer in its instrument’s limit order book on the venue, depending on the order’s side (buy or sell). If the order is then found to be taking prices (i.e., buying at or above the prevailing best offer, or selling at or below the prevailing best bid), it is put in either the ‘taker buying’ or ‘taker selling’ buffer for its instrument. Otherwise, the order is making prices and is put in the ‘bid at P ’ or ‘offer at P ’ buffer for its instrument, where P is the limit price of the order. In this way there are a plurality of different buffers for each instrument traded on the venue—notably one for each resource that may become in contention. These resources are, of course, queue position at each price level in the limit order book for price-making orders, and standing bids and offers in the limit order book for price-taking orders. As for the timer in the event-based approach described earlier, if the buffer is empty the first order assigned to it ‘is’ the event, and therefore starts that timer. Subsequent (competing) orders are added to this buffer until it reaches T , which on TRM is set to 3ms. Upon the timer reaching 3ms the orders in the buffer are shuffled and drained from it, ultimately being presented for matching on the venue in a generally different ordering to which the buffer received them. The procedure by which the shuffling and draining occurs is as follows. First the orders are ‘bucketed’ by participant, then that list of participants resulting from the bucketing is randomized such that each participant has an equal probability of appearing before (or equivalently, after) any other participant in it. The randomized list of participants is repeatedly iterated over, from first to last participant, and at each step only a single order for a participant is removed before moving onto the next participant. Where there are multiple orders for a participant in the buffer, the oldest remaining one (i.e., that which was received by the buffer earliest) is that chosen for removal. The repeated iterations over the list of participants, of course, continue until the buffer is empty.1 To complete the description of the refinement on TRM some other details also warrant mention. Unlike on other electronic trading venues, where due to marketmaking activities it is likely the most widely-used message type, the cancel/replace1 The

pseudocode for this shuffling and draining procedure is provided elsewhere [Melton 2015, par.59-68]. Under review, May 2017

4

·

H. Melton

request operation is not supported on TRM. This is fortunate because ‘shoehorning’ this operation into approach described above, although possible, is also cumbersome [Melton 2015]. Cancel-order operations are, however, supported on TRM and these continue to be processed in real-time by the venue, i.e., are not subject to buffering. Further discussion of this is provided in Section 3, but in short this decision seeks to address the predator-prey relationship that might otherwise exist between a fast taker and slow maker on the venue [Farmer and Skouras 2012]. 3.

CONSEQUENCES OF THE DESIGN

What is arguably the main innovation of the design—its ability to detect and distinguish individual races that occur for different types of resources on a venue—is also the thing with the most interesting consequences. In particular, by detecting the start of a race through the receipt of the first order in it, and by starting its buffer’s timer upon the receipt of this first order, the length of the delay imposed by the mechanism is roughly equal to the sum of the minimum response time sought to be established for participants and the magnitude of the temporal unfairness it solves for on the venue [Melton 2017]. This is not true of other mechanisms that have appeared in the literature [Harris 2013; Budish et al. 2015]—they can only establish short minimum response times through the imposition of very long delays, and may not be able to solve for temporal unfairness at all [Melton 2017]. Good reasons for seeking to minimize the delay imposed by a venue in its processing of orders exist and are described elsewhere [Melton 2017]. Another interesting consequence of the design’s ability to detect and distinguish individual races, is that a participant cannot ‘artificially’ start the timer for one race through the submission of an order that belongs to a different race. For instance, if there were instead only one buffer per instrument (and not one buffer per resource type) then a fast participant could, with no intention to be matched (and little risk of it actually occurring), submit a very low-priced bid to the venue so as to start the instrument’s timer early ahead of a forthcoming event whose publication time is known in advance. Then, responsive to the observation of this event, and just before the timer had expired, the participant could then send an order to take liquidity from the book—an act which may provide large economic reward to the participant—thereby disestablishing the minimum response time for that event the operator of the venue sought to ensure. The resistance of the actual design to this form of manipulation by participants seems reminiscent of strategy-proofness in game theory. Yet another consequence of detecting and distinguishing individual races is that participants who submit a plurality of different orders ‘all at once’ are not disadvantaged relative to other participants who submit relatively few orders at a time. Again, if instead there were only one buffer per instrument, a market-making participant that is simultaneously submitting bids and offers on the instrument may be disadvantaged relative to two other participants that are both trading more directionally by each submitting only a single bid and offer, respectively, on alternate sides of its book. By virtue of this single buffer and the procedure for draining it where only a single order is removed per participant per iteration, the marketmaking participant is afforded an equal chance of winning the race on one side of Under review, May 2017

Market Mechanism Refinement on a Continuous Limit Order Book Venue

·

5

the book with one of the directional traders, but disadvantageously is guaranteed to lose the race with the other directional trader on the other side of the book. Thus by having a plurality of buffers per instrument, and by grouping only competing orders into each buffer, a participant has an equal chance of winning each individual race on the venue even when s/he simultaneously competes in a plurality of such races. This is important because venue operators tend to value marketmakers highly; ensuring they are not discouraged or disadvantaged in performing this activity is thus an important consequence of the design. A final consequence of detecting and distinguishing individual races is the manner in which it informed setting of the delay period. The 3ms delay period on TRM is long enough to both solve for temporal unfairness and to establish a minimum response time, but perhaps surprisingly is much smaller than the maximum time difference observed in order processing times among competing orders and separately in market data distribution to different participants. The insight this aspect of the design eventually led to was that different race types (e.g., racing to take the bid or offer, racing to establish a new top-of-book price, or racing to join an existing top-of-book price, and so on) have different profit-and-loss functions. As noted of skips and slips in the literature on queue theory [Gordon 1987], the benefit a participant receives from ‘overtaking’ another in a race may statistically be offset by the penalty imposed on that participant by later being ‘overtaken’ in that same type of race. A method was subsequently devised for empirically determining the minimum delay the mechanism would need to impose on a venue so as equalize overtaking and being overtaken among all pairs of participants in each race type [Melton 2016b]. Advantageously, applying this method on TRM led to a delay period that both ensured temporal fairness and established a minimum response time, but was much shorter than the extreme time differences it and other distributed systems tend to exhibit as ‘tail latency’ [Li et al. 2014]. In terms of the design’s consequences on other aspects of participants trading strategies, consider the following. Market-making participants may sometimes send a plurality of small orders ‘all at once’ at the same price-level in the book. It is important for them to know which were inserted nearer the front of the queue at the price level, and which were inserted nearer the back, so when they want to cancel some of those orders, they can cancel the ones nearer the back first. Maintaining the temporal ordering of each participant’s orders in a race is thus important, and is why a participant’s orders are drained from the buffer from oldest-to-newest. Another reason for this relates to taker orders that ‘sweep’ the book. If the following sequence of taker orders (which mirror the state of the ask book) are sent by a participant ‘all at once’: buy 1Mio at 1.00, buy 1Mio at 1.01, buy 1Mio at 1.02, they will only all match if they are processed in that exact sequence. If they are not, and say the buy 1Mio at 1.02 order is processed first, it will match against the 1.00 offer and consequently it will not be possible to match the buy 1Mio at 1.00 order. Regarding the ‘asymmetry’ in how the design treats taker and maker buffers—the latter grouping by orders’ limit prices—consider the following. A market participant expecting to receive price improvement on his/her taker order sends an order to buy at 1.01 when the prevailing best offer is 1.00. In the case that participant ‘misses’ Under review, May 2017

6

·

H. Melton

the 1.00 offer and thus does not receive that price improvement, the order will (hopefully) then immediately fill at the worse price of 1.01 saving the participant from sending a second message and incurring the cost of another round-trip to the venue in doing so. By having taker buffers that do not group by limit price this participant’s strategy is able to remain unchanged in that s/he is not suddenly disadvantaged when competing with a faster participant that might send an order to buy at 1.00, because both orders will be put in the same ‘taker buying’ buffer. To briefly address some other consequences of the design—trivially, bucketing by participant eliminates the incentives that might otherwise have existed for participants to send multiple redundant orders to a venue [Budish et al. 2015, p.1611]. Next, because most orders on TRM are not in competition with others they are subject to exactly a 3ms delay which participants may prefer over variable delays [Melton 2017], and which may eliminate incentives to ‘time’ the sending of orders to a venue so as to minimize such delays, e.g., just before a batch auction window closes [Budish et al. 2015]. Finally, the single order per participant draining procedure may further defend against manipulation of the minimum response time by helping to ensure that if a fast participant say buys a small quantity in anticipation of a forthcoming event, his/her subsequent attempt to buy a larger quantity upon actual observation of that event will be undermined by a similarly fast participant reacting to the event itself and also buying. Finally, the consequence of the decision not to subject order-cancel requests to buffering is that a maker is guaranteed to be able to cancel his/her order whenever the cancel request is processed by the venue within 3ms of the first order seeking to take it. In this sense, a minimum response time is established for makers to be able to cancel their orders ahead of the fastest takers, and the previously mentioned predator-prey relationship that might otherwise have manifested is eliminated. Other approaches to treating cancel requests more like taker orders are described elsewhere [Melton 2015], but all are more cumbersome—and indeed may lead to new forms of manipulation—than simply not subjecting them to buffering at all. 4.

CONCLUSION

The refined market mechanism described in this letter has been successful in that (1) it has been well-received by participants on TRM, and (2) there are no plans to remove it from the venue or to further modify it. While deemphasizing speed is becoming increasingly common in FX [Detrixhe 2016], it remains to be seen the extent to which it will take hold on major venues in other asset classes. Finally, most of the approaches for deemphasizing speed that have been proposed so far and that are in-use today are forms of the general technique of buffering; the extent to which non-buffering approaches [Melton 2016a; Kyle and Lee 2017; OneChronos 2016] will come into actual use also remains to be seen. REFERENCES Budish, E., Cramton, P., and Shim, J. 2015. The high-frequency trading arms race: Frequent batch auctions as a market design response. The Quarterly Journal of Economics 130, 4 (November). Detrixhe, J. 2016. Speediest traders becoming less welcome in curUnder review, May 2017

Market Mechanism Refinement on a Continuous Limit Order Book Venue

·

7

rency markets. https://www.bloomberg.com/news/articles/2016-12-11/ speediest-traders-are-becoming-less-welcome-in-currency-markets. In Bloomberg News, accessed Jan 9 2017. Farmer, D. and Skouras, S. 2012. Review of the benefits of a continuous market vs. randomised stop auctions and of alternative priority rules (policy options 7 and 12). https://goo.gl/ QDpgSP. UK Government Office for Science, accessed Jan 7 2017. Gordon, E. S. 1987. New problems in queues: Social injustice and server production management. Ph.D. thesis, MIT. Harris, L. 2013. What to do about high-frequency trading. Financial Analysts Journal 69, 2. Hasbrouck, J. and Levich, R. M. 2017. FX market metrics: New findings based on CLS bank settlement data. Tech. rep., National Bureau of Economic Research. Kyle, A. S. and Lee, J. 2017. Toward a fully continuous exchange. Available at SSRN 2924640 . Li, J., Sharma, N. K., Ports, D. R., and Gribble, S. D. 2014. Tales of the tail: Hardware, OS, and application-level sources of tail latency. In Proceedings of the ACM Symposium on Cloud Computing. ACM, 1–14. Melton, H. 2015. Ideal latency floor. US Patent App. 14/533,543. Melton, H. 2016a. Systems and methods for obtaining and executing computer code specified by Code Orders in an electronic trading venue. US Patent App. 15/064,163. Melton, H. 2016b. Systems and methods for quantifying temporal fairness on electronic trading venues. US Patent App. 14/930,499. Melton, H. 2017. Understanding and improving temporal fairness on an electronic trading venue. To appear in The Sixteenth International Workshop on Assurance in Distributed Systems and Networks (ADSN2017). OneChronos 2016. OneChronos website. https://www.onechronos.com/. Accessed May 29 2017. Phelps, S., McBurney, P., and Parsons, S. 2010. Evolutionary mechanism design: a review. Autonomous Agents and Multi-Agent Systems 21, 2, 237–264. Simon, H. A. 1996. The sciences of the artificial. MIT press.

Under review, May 2017