Ciamac Moallemi: High-Frequency Trading and Market Microstructure

I thought part of the purpose of this is to sort of maybe acquaint people with research of faculty in general so I'll say a couple of words about things I work on in general before I get into this topic I'm really an applied mathematician and about half my time I spend thinking about stochastic control problems these are problems where you have to make decisions over time and there's uncertainty about what's going to happen and they're really you know a class of abstract mathematical problems that are very challenging and fundamental in the sense that many you know problems in engineering or in business and so on have that type of flavor so that's about what I do with half of my time prior to actually becoming an academic I did a number of things in the working world and one of those was I manage a hedge fund for a number of years and motivated a little bit by that the other half of my research is really the more applied side and I think about stochastic control problems in the context of financial engineering and specifically the kind of problems I'm interested in our problems are surrounding optimal trading market microstructure high-frequency trading and so on and so forth so that's what I'm gonna talk about today so um what I'm going to talk about is give you a little bit of flavor of the issues in the you know modern electronic markets and in order to sort of give you some appreciation for that you know I think it's worth maybe going through what are some of the main features of US equity markets today because really they're quite different than they were five and certainly ten years ago so so what are some of those features first markets are predominantly electronic right so you see this backdrop on CNBC where there's the New York Stock Exchange and people are yelling at each other and so on and so forth to a large extent that is irrelevant right that is not where the UH the trading happens the trading happens on computers and electronic trading has really certainly inequities dominates as the a the primary mechanism of exchange the second and this is a little bit counterintuitive we like to think of something like an exchange as a mechanism for centralizing trade of bringing buyers and sellers together so that you know there aren't search frictions and so on and so forth well you know really in the five years what's happened in the u.s. is that the opposite trend has occurred is a trading has become decentralized or fragmented in particular for um for various reasons there's no longer let's say one primary exchange it used to be that for any particular stock you know is predominantly traded either on Nasdaq or NYC but now there's you know a handful of them and they're all important in the sense that each of those you know exchanges you know accounts for at least maybe 5% of equity trading so there's many venues trade is no longer centralized most of the venues are organized as exchanges they account for about 70 percent of trade and these exchanges are operated typically as electronic limit order books in the sense of it's a open market people can submit orders to buy and sell and they attach prices and when the you know prices cross there's trade and so on this is as opposed to a dealer market or a specialist market which is the way historically the New York Stock Exchange was organized however that mom would life would be too easy if we had just um one kind of market structure so about 30 percent of trade occurs on alternative kinds of venues right so here we have things like electronic crossing networks dark pools internalization so on and so forth and finally maybe from from my perspective the most striking feature is that the participants are increasingly automated it used to be that if you were at say a hedge fund and you had a portfolio manager and he wanted to buy a million dollars worth of Google there's some guy who's a trader and he knows how sort of these things works and he you know makes calls and does stuff and you know it happens over time right well now computers do that right so on the buy side investors under the rubric of algorithmic trading either themselves or on an agency basis as a service provided by brokers they'll take large paranor orders and slice and dice them over time and across exchanges and so on and so forth and trade them similarly the guys who are providing liquidity of the market makers again these used to be sort of human traders now in most of these markets oftentimes they go under the rubric of high-frequency trading one dominant kind of high-frequency trading is essentially providing liquidity providing market making services so overall the you know these are all quite recent trends this last one is particularly great from from from my perspective because as someone who likes to solve quantitative decision-making problems that means there is a stuff for me to to do here but you know the overall effect of this is that the world is much more complicated than it was before right and the interactions between you know an algorithmic trader and a high-frequency traders and so on are difficult to predict and in fact you get things like this right so this um picture probably you've seen this is the famous flash crash of May 6 2010 and you know the sec did a report on this basically what happened is in about five minutes the market fell 5 percent based on no news or fundamental and information or whatever and then the next five minutes it sort of recovered so it's kind of a blip that you know according to this report at least came about from some kind of pathological interaction between a large algorithmic trader and high-frequency traders right so this sort of raises in my mind I think two classes of important questions right one from the perspective of you know the system from the perspective of policymakers regulators and so on is this complex structure that we have is it good do we want things like you know dark pools is it is it good to have so many exchanges and so on and so forth right how do we avoid something like this right so those are the the questions that a positive policy level but even at the level of individual participants we still have to solve these decision problems right though the world is out there for what it is now and if I'm trying to you know buy some stock I have to decide am I going to use the dark pool am I going to use in exchange how am I gonna do I'm accomplish this so I'm interested in both of these classes of questions and what I'm going to talk to you about today is two specific problems related to high-frequency trading and market microstructure the first is going to be understanding the importance of latency and the second is going to be understanding of the role of dark pools in markets right and so these are two two different research projects I've worked on recently so let me start out with latency what do I mean by latency roughly speaking I mean the the delay between making a trading decision and its implementation right so if I decide to a hundred shares of Google and I transmitted that order to Nasdaq how long before you know that quantity is taken from the the order book or similarly maybe I have an order outstanding and I want to cancel it how long does it take between when I make that decision and when that one order is pulled from the matching engine and is no longer eligible for a execution this sounds like a little bit of an arcane topic and I think it was right used to be the domain of you know maybe IT people and and so on and so forth but really in the past few years it's entered sort of the public discussion and even my mom could tell you about latency because we see articles like this right so this is an article that really sparked a lot of controversy when it came out a few years ago now it was in the New York Times talking about stock traders find speed pays in milliseconds so let's see the top talks about you know powerful computers our house next to the machines that drive market places that's the idea of colocation will come back to high-frequency traders often confound other investors by issuing and cancelling simultaneously so you know maybe somehow it's good to be able to trade kind of quickly although you know bullying doesn't sound so good here's a quote from someone from New York Stock Exchange it's a technological arms race what separates winners and losers is how fast they can growth right so clearly a an important discussion and and raises a question is this a good thing that we're able to trade so fast right now one kind of issue is that when you see sort of quotes like this to a large extent the the the conversation is dominated by people with an interest right so for example if you happen to be running one of the laggards in terms of technology in financial exchanges maybe you have one view if you're a high-frequency traders maybe you have another view right so you know I think our role as an academic is to sort of dissect this and really to build models to understand what these you know phenomena represent and a question like this is being able to trade with very low latency with very very high frequency is is that good or bad I'd like to approach that question but I'm going to ask them even easier question to start with why is latency important right clearly if people are willing to spend a lot of money to go out to sort of achieve that what is it worth and how can we and that so that's the question I want to answer I want to value the importance of low latency and phrased a different way what is the cost associated with having latency what is the cost associated with being a slowpoke okay so that's what I'm going to talk about to give you a little bit of a perspective of what the numbers look like this is the evolution of latency in US equity markets over the the past 35 years prior to 1980 it used to be that if you put in an order to buy a stock it would take two minutes for that to actually occur of course it's a timing people make phone calls and stuff and that sounds ridiculous from our perspective today but put that in a little bit of perspective in you know that type of era in a typical stock there was a trade every 20 minutes so if a trade takes you know 2 minutes but you know there's only 1 of them every 20 minutes maybe that's not that's not so bad in any event that's what it was there was an event in 1980 where they I think they installed a big mainframe and they were sort of very proud at how the their the processing times came down and it came down to about 20 seconds of course if you sort of fast-forward to a current times you know 2007 to go kind of latency numbers in the hundreds of milliseconds and to give you a sense of what that means if you ask a human to perform the most basic task like if I flash a light into your eye and ask you to press a button when you see that that takes a couple hundred milliseconds right just reaction time right so if you're actually making trading decisions on that time scale it's not humans it's computers trading with each other right go forward another couple of years now you're in the single-digit millisecond right and here physics and the speed of light starts to become important delays for propagation of information are significant right so if you want to send a piece of news from Chicago to to to New York you know speed of light limits you that can't happen faster than you know I don't know what it's for five milliseconds right so if you want to be trading in less than a millisecond that means you need physical proximity right and that's the idea of colocation so not only are humans out of the game it's a bunch of computers trading with each other but essentially a bunch of computers and a handful of data centers in New Jersey right so and you know the trend is to even go below that so it's state-of-the-art right now is four co-located high frequency trading is you know round-trip times and the hundreds of microseconds or less right so this has been driven by a number of things one is technology we can do this in 1980s you know they didn't have the computers that were and the networks that were able to do this but but the second thing is there's demand people are willing to pay for it right part of the the way that all these different exchanges emerged is they offer technology benefits relative to the incumbents right so you know why might it be important to be able to trade very quickly right and to who might not be important is it only important to high-frequency traders should pension funds care should retail traders care well some of it depends a little bit on who you are and what you're doing but we can come up with a couple of hypotheses one is that you know typically you know it's best to make decisions with the latest information possible right so for example if you're you know out there looking to you know I don't know sell a hundred shares of Apple stop maybe the price that you're willing to sell out depends on a bunch of things right it depends on the price other people are willing to sell or buy out depends on the price on other exchanges depends on maybe what Microsoft stock is doing so on and so forth so as you digest more recent information maybe that will alter the UH the price you are willing to sell at and so there would be some advantage to having low latency right this is really latency in an absolute sense another thing you might think about is well maybe the absolute latency isn't so important but the relative latency right if you have two traders and they're doing sort of very kind of similar trading strategies typically the flavor of these things is that the winner takes all so the fastest guy will get all the profits and the other guy will get knocked out and so you know so they're you know it's not really important that you trade in under a millisecond it's just important that you're faster than the next guy right and a third effect is the rules by which exchanges are organized there is a priority for being early so there's some advantage of being early you show up at the beginning of at the top of the order book and that offers certain advantages and so on so depending on who you are these different effects might kick in and what I'm going to talk about is a model that really addresses the first one right how does investor benefit from having access to sort of the latest information in terms of lowering their their costs so here's a model right we have a stylized execution problem we have a trader who wishes to sell a hundred shares over a very short time horizon let's say ten seconds and this is meant to be a problem that you know every trader faces at one level or another and as I was describing to earlier the the value that the trader perceives it evolves over time and the the price of which the trader wants to get depends on this value right so in order to figure out how best to sell this hundred shares the trader has to observe this value process and if we add a little bit of latency that introduces a tracking error right so now the trader can't does it precisely know what the value is you only know the value you know maybe a millisecond ago and because of that they have he has to alter his actions and that creates a cost so latency becomes a friction and what we do is we quantify the on the the cost associated with latency if I look at this execution problem I look at the transaction cost in the presence of latency see how much worse that is and if you had known a latency and normalize right and what we got is we get a very simple expression right the the particular formula here isn't important what is important is it depends on commonly observed them on market parameters it depends on the volatility of the stock right the more volatile the stock is a more important latency is and it depends on the other bid-ask spread right the more liquid the stock is the more important latency is and instead of looking at the formula we can look at a picture here's what the picture looks like I took that formula and I calibrated it just for someone who's trading goldman sachs on that particular day and we can see as we go from the human time scale of let's say 500 milliseconds really fast human to the Machine time scale of a millisecond what we see that latency goes from being about 20% of a transaction cost to being you know of 1 or 2% right so does this make sense is this significant well let's try and interpret that so let's imagine we took the stock and we normalized it so that the bid offer was a penny most stocks in the US have a bid offer of a penny right what that thing is suggesting the the plot I showed you is that the the value of decreasing latency from the human time scale to the Machine time scale is about 20% of a penny or 20 mils right that seems like a very small number but you want to guess how much high-frequency traders make people who've made the investment to be able to trade on this kind of time scale well nobody really knows but um this kind of self reported numbers are of the same order of magnitude right a high-frequency trader you know again based on these sources might typically expect to make in the tens of million right very very very kind of a small margin same order of magnitude similarly let's say you didn't have the ability to to trade electronically and you wanted to farm it out you wanted to pay an investment bank to trade for you and presumably they've made that investment right how much would they charge you they charge you a similar kind of number right so typical fee for algorithmic trading execution would be on the order of of the same order of magnitude right and it's of the same scale of the cost so this tells us a couple of things right number one I'm latency is potentially important to all investors right so I wasn't you know this is the basic problem I started with the promise showing a hundred shares in in ten seconds everybody has that problem retail guy might have that problem if you're you know really trying to sell a million shares well you know it's some kind of fine time scale that breaks down into individual trades of a hundred shares so it's important to everyone right but how important it is depends on what the rest of your costs are if you're the most at the most efficient cost level in terms of the Commission's you've negotiated and so on and so forth latency is worth about as much on the other hand if you're a retail investor you're not paying five mils per share traded you're paying ten dollars to each way that's orders of magnitude more than any of this right so from my perspective for for a retail investor this kind of stuff doesn't matter and the commissions and other things you're paying way dominate the UM though the value of latency okay so some cross-sectional stuff but I won't talk about that the second topic I'm going to um briefly mention here is um dark pools dark pools are an alternative trading mechanism and the name sounds kind of foreboding but the basic idea is is simple if we think about something like a limit order book right if I want to buy typically there is an offered price which is higher than the price at which I could saw which is a bid price right so there's a bid offer spread and in you know again something like a limit order book with an exchange someone is providing liquidity right maybe a high-frequency trader and you know they're gonna charge for that and this bid-offer spread is what they charge right now um what's an alternative mechanism the idea of a dark pool is instead of having you know these intermediaries posting orders let's let people directly trade with each other so let's just have an anonymous pool where people will you know some people can declare they want to buy some people can declare they want to sell and if there's a match they'll be matched with each other and it will occur at mid market so no transaction costs right so that sort of sounds good on paper but the key is you're giving up something right if you try and buy it on an exchange with a market order you're gonna execute for sure right if you put an order into a dark pool you'll get it at a better price but you may or may not get it right and this is actually I think a trade-off that occurs in many markets right a trade-off between uncertain trade at a better price ie the dark pool or guaranteed trade at a worse price right so for example an eBay right and an eBay auction you typically buy at all you can pay us a price premium get the UM the item you want with certainty or you can participate in the auction maybe you'll get it for cheaper maybe you won't get it online advertising the way display ads work or so on there's a couple of different pricing models right a publisher can charge a advertiser based on just showing the impression based on you know only if people click on the expression maybe based only if there's a transaction right so here you have different degrees of certainty but of course if the you know you're going to charge by transaction you're going to charge way way more than if you charge just by a click and we can see financial examples here what I'm going to talk about here is a simple stylized model in a financial context where investors have two options one is what I'll call a guaranteed market where you can trade with certainty but you pay a transaction cost you pay the bid-offer spread right and so here I'm thinking of something like a dealer market or electronic in an order book or something like that and the second option is a dark pool here you put your order into one of these electronic crossing networks if trade occurs it occurs in mid market there's zero transaction cost but you're not sure it's going to happen right so we're trying to evaluate these two alternatives we have a model and the key ingredient in this model is information ladders a lot and this is a ties in a little bit with what Ray was talking about in industry parlance this is a short term alpha and what I mean by information is do you have any information about what's going to happen to the price over the short term over the time horizon of your trade being executed right this might be important for um for for two reasons number one if I think the stock is going up right and let's say I'm in fact I'm pretty sure that the stock is going up that's gonna affect whether you know I want to trade with certainty or I want to be uncertain right for example err if I'm certain maybe I'm willing to pay the transaction cost and I want to trade with certainty so my information clearly matters right however others people's information matters also right why because I'm when you trade you're trading with others in the in the case of a dark pool and if you're systematically trading with people who have more information than you maybe that's not going to work out for you so well in the end right so one critical thing is going to be modeling information we're going to have a model where we have three kinds of traders we have speculators right everybody observes some kind of signal about what's gonna happen to the price the speculators they're just trying to make money off the price swings right on the other hand we have intrinsic buyers and sellers they also would like to buy low and sell high however they they have their own reasons to trade as well they have idiosyncratic a kind of desire to trade maybe your pension fund and your rebalancing your portfolio so you know I don't know you want to buy that Google stock cheap but you know even if you have to pay a little bit more for it you're still willing to buy it because you you know you get utility from from from from having it we solve for an equilibrium in this model and we get a number of the interesting predictions so number one who chooses which market place well as you might expect the more information you have the more you're going to go to the guaranteed market place if you're very well-informed and you know the price is going to go up you want to buy with certainty on the other hand if you less informed if you have little or no information maybe you're only trading for four idiosyncratic reasons well you know you're willing to trade in the dark ball you're willing to take that risk because it doesn't make sense for you to pay transaction costs right now that sounds reasonable but that has a couple of implications one is if we imagine the world with the dark pool and the world without the dark pool transaction costs will be higher in the presence of the dark pool why is that where do these transaction costs coming from these are these are bid-offer spread is going to be set by market makers who are trying to make money if you have a dark pool present and it is systematically sorting the the traders right and so only informed traders go and trade in the guaranteed market well those market makers are gonna end up systematically losing more money and so what are they going to do they're going to widen spreads right in order to compensate for that right so the presence of the dark pool is actually going to deteriorate the quality of the the guaranteed market investors in the dark pool are going to experience adverse selection what do I mean by that if you trade in the guaranteed market you know no matter whether the price is going to go up or down you're going to get that share right if you trade in the dark pool you know you're not sure if you will or not but if you're trying to buy what will happen is typically when the market is going down your order will get filled and when the market is going up it won't so precisely in the circumstances where you don't want to trade because you could have bought it cheap later you will trade and otherwise you won't this is something that I think is very crucial that people don't understand in practice or a surprising number of people right so an IE person will look at dark pools and say there's no transaction fees you're trading at mid market well because you're not trading for sure and because your trades are going to be correlated with what happens to the price afterwards you are in fact paying a type of adverse selection fee it's just implicit it's not explicitly the bid-offer spread you see the bid-offer spread it's you know clear what you're paying here statistically you're paying this fee but it's very real and it can be of the same order of magnitude as a bid offer spread right so the dark pool isn't as good as it looks and finally maybe what's a most interesting in my mind overall introducing the dark pool decreases welfare right so what do I mean by that let's try and think about whether this is sort of good or not we have a bunch of shares floating around in a bunch of money if I think of transfers of money from people in different traders in the system that sort of doesn't matter that's zero-sum if I buy the stock from you and I pay you know a dollar or less you know maybe I gain a dollar but maybe you lose a dollar from the perspective of you know the system it doesn't matter right on the other hand the shares do matter right why because people have different intrinsic value for those right there's this there's people who have a lot of value to owning them and people who don't right so what you would ideally like is you'd like these intrinsic sellers to be selling to these intrinsic buyers in order to maximize the utility in the system and that's the concept of the this I'm social welfare right well it turns out in the presence of the dark pool impedes that and reduces the welfare of the system so I'm over time so I'll stop here

Channel: Columbia Business School

Views: 20,230

Rating: 4.9228296 out of 5

Keywords: High-Frequency Trading, Market Microstructure, Ciamac Moallemi, Asset Pricing, Columbia Business School

Id: y4z1FSdN_GA

Length: 25min 17sec (1517 seconds)

Published: Mon Nov 19 2012