Tuesday, February 11, 2014

The HFT arms race

High Frequency Trading costs real resources. It consumes computing power and the mental effort of smart people. When we decide if HFT benefits the world, we have to weigh these costs against whatever benefits HFT provides.

What are the benefits of HFT? The normally cited benefit is that HFT "increases liquidity". And indeed, since the introduction of HFT, some types of trading costs - commissions and fees - have gone way down. If HFT reduces the total amount that America spends on trading assets without reducing the efficiency of the market, then HFT has created value, by replacing something expensive (brokerages and dealers) with something cheap.

But what if HFT consumes liquidity instead of increasing it? Theory suggests that if HFT consists of a bunch of algorithms trying increasingly hard to beat each other to the punch, then liquidity will go down, and the resources spent on HFT will just be a waste. Now, via Johannes Breckenfelder of Stockholm's Institute for Financial Research, we have evidence to back up the theory.

Unlike other countries, Sweden's government allows some finance researchers to observe the identities of traders, so we profs can see who is doing what. That's hugely important, because in most data sets, we can't tell which trades are submitted by HFTs. Now we can. Breckenfelder, after going through the arduous process of obtaining this amazing data set, has begun to sift through it for insights into the inner workings of the market. His first big discovery is that when HFTs trade against non-HFTs, they increase liquidity, but when they trade against each other, they end up removing liquidity from the market. From the conclusion of Breckenfelder's paper:
High-frequency traders (HFTs) play a role of critical importance for the financial markets. HFTs exploit not only liquidity-providing short-term investment strategies (e.g., market making), but also liquidity-consuming short-term investment strategies (e.g., directional trading). When HFTs face competition from other HFTs, liquidity-providing strategies will improve market quality, while liquidity-consuming strategies will naturally worsen market quality. We find that competition among HFTs coincides with a decline in liquidity and an increase in liquidity-consuming high-frequency trades as well as in high-frequency momentum trading.
Note that one assumption in this result is that directional trading consumes liquidity. This is probably a good guess, since a pretty simple HFT strategy is to use someone else's order as a signal - if you see someone else buy 10 shares, for example, it's a good bet they're going to buy 10 more, so you buy in order to take advantage of the price rise that their next order will generate. That obviously makes it harder for the other person to trade, and hence decreases the liquidity of the market. But if prices have momentum for other reasons, it's actually conceivable that directional trading might increase liquidity, through processes that are poorly understood.

Does Breckenfelder's result mean that we should curb HFT, for example by changing the current market system into one involving batch auctions? That depends on several things. Liquidity is not the only factor in market quality - there is also informational efficiency, or how well prices reflect available information. The impact of HFT on informational efficiency is not well-understood, because it depends on adverse selection, whose action in financial markets is not well-understood. Another crucial piece we need to understand is corporate finance, or how liquidity and efficiency at various time horizons distort firms' capital budgeting decisions. That's a huge area of research, of course, and as far as I'm aware, no one really knows how much market quality at high frequencies affects the economic efficiency of corporate investment decisions. A third thing we need to know is HFTs' effects on market stability, which is something people at Stony Brook are working on (nor do we know the true economic cost of things like "flash crashes"). So it's very difficult to compare HFT's costs and benefits at this point.

But anyway, Breckenfelder's result implies another hypothesis: As the market gets saturated with HFTs, liquidity will bottom out and start to creep back up, since HFTs will be trading more with each other and less with non-HFTs. There is evidence that this is happening. Profit at HFTs has plummeted and trading costs may have begun to creep back up. HFT, it turns out, might be a small, self-limiting phenomenon. The question of whether it is valuable is still very interesting, of course, but the size of the sector might top out at so small a value that regulators shouldn't lose too much sleep over it.


John Cochrane flags another interesting paper on the topic, whose title is the same as the title of this blog post (the paper obviously came first)...

Zachary David, an HFT trader himself, has some good skeptical comments on Breckenfelder's paper.


  1. The cost of a trade also includes the bid-ask spread, since we assume the "true" price is somewhere in between. If we increase the minimum increment back to 5 or 10 cents, there will be more liquidity at every level and also less incentive for machines to pull orders (due to losing queue priority).

    In practice we see this happened just a couple years ago. ICAP's EBS currency platform moved almost all of their minimum increments to decipips (0.00001), and as a result liquidity dried up as market makers didn't want to deal with adverse selection. At the end of the year they decided to return to full pip increments and indeed, liquidity came back!

    It wasn't long ago that equities markets also were in minimum increments of 5 and 10 cents (or more). If we want to provide for continuous price discovery as well as increased liquidity, the most rational choice is to move to wider increments. (periodic auctions are nonsensical)

    1. Also, on a point of mechanics. The social and institutional cost of implementing periodic auctions would be tremendous. There is decades of legacy code that would have to be overhauled at every level. Everyone from your grandpa to Johnny Fastmoney would be affected.

      But increasing the minimum price increment is simple. Every exchange already sends out what is called a "Security Definition" for each product that includes this information already.

    2. Eliminate fractions, that'll do it. Minimum one penny diff between bid/ask, nothing executes in the gap. These HFT's exploit really tiny diffs.

    3. Paul the fractions come from dark pools.

  2. Anonymous6:02 PM

    HFT's primary benefit is that it is financing chip and system development; nothing to do with finance.

    1. Sounds like the argument for porn's benefits on development of internet infrastructure.

  3. I think we should really look into the regulation of blogging. I mean, look how many smart people use up all their time to blog, time that could be spent making prosthetics for toddlers who lose their legs? And think about how many energy we could save by axing all those servers that host blogs!!! That's way more energy than HFT could ever hope to consume.

    The usual argument in favor of blogging is that information exchange helps improves peoples ideas. But really, though, how many people actually change their minds based on arguments in blogs? I'd say fewer than 5% of blog readers have ever actually changed sides on an issue based on arguments they read on a blog.

    1. Anonymous8:08 AM

      "And think about how many energy we could save by axing all those servers that host blogs!!! That's way more energy than HFT could ever hope to consume."

      I am 100% sure it's really true that HFT consumes less energy than blogs. Nowadays hosting blogs does not take too much computing power, while HFT probably does. I know an algorithmic trading company (low frequency), they switched to GPUs more than ten years ago, before it was fashionable.

  4. Anonymous9:06 PM

    The two easiest ways to curb HFT abuse are:

    1) Force all resting orders to exist for at least a couple of milliseconds to allow other traders to actually trade on them. (This prevents players from putting misleading quotes inside the true spread that are not able to be traded upon by others.)

    2) Charge the HFT traders per quote and rebate them per actual trade, to diminish practices such as quote stuffing.

    Abuses exist in the exchanges because traders do not have to pay for the costs they impose on others.


    1. All of those things already exist in certain markets. Some exchanges have minimum order life parameters. Some exchanges have ticket charges. Some exchanges have messaging limits where you have to have a certain ratio of trades to messages.

      The reality is that this entire discussion is a non-issue. The paper in question didn't find anything that people actually making trades haven't known for years.

      The market already has a number of solutions: large trades (blocks) are done on dark pools, through brokers, or through smart execution algorithms.

      Liquidity is autocorrelated. A market maker does not want to be the only order at a certain price, because once that order is filled the market is already going against them.

      I see little point in forcing an exchange to abide by some type of clearing mechanism or specific order handling rules. People will trade on the markets they think they have the biggest edge on. There are tons of options already.

  5. As long as they're abiding by the rules of the exchange, there is no such thing as HFT "abuse".

    1. Anonymous9:47 PM

      Well that's definitionally true for your definition of "abuse" and "rules", I suppose. But not helpful to the conversation.

    2. Well, how do you normally define "abuse"? If abuse is the misuse of something, or using something in a different way than it was intended to harmful effect, then what would that mean for trading on a regulated exchange? It seems like the main functions of an exchange are to facilitate 1) the transfer of risk between between participants (mostly via speculators), and 2) price discovery.

      It's hard to see how HFT games could interfere with 1. Grandma isn't trying to compete with HFT firms in her retirement account, and neither is ADM with their corn hedges.

      I also don't see how HFT could do anything to interfere with 2. The idea is you set up some transparent, uniform rules, and then you let profit-motivated people place their bets according to whatever info they can get their hands on, eventually reflecting all that information in the price. As long as players are putting their money on the line and everyone knows the rules, how can their activities not aid price discovery?

      I mean, what kinds of things do people bitch about HFT players doing? They do stuff really fast. OK. You might argue that the price discovery benefit is really small given the costs to get fast, but that's not the same as arguing that it doesn't help or harms price discovery. Transferring information quickly is better. That's what the AT&T commercials have taught me.

      The other primary complaint is that some HFT players might "spoof" or otherwise try to temporarily manipulate the market price artificially. Even this, though, reflects important information. To make money this way, you have to scare more "normal" market markers into doing something dumb, like puking their position to you at a bad price. What would make them do this? The only reasonable answer is a lack of confidence in their pricing, which means they didn't have enough information to justify markets as tight as the ones they were making before being hit. If they subsequently make wider markets, then the new, thinner market more accurately reflects how much info is actually available about that securities price.

  6. There's another cost to HFT that is perhaps large. Fear of being front-run is driving much of market trading into "dark pools." There are efficiency and regulatory concerns with this; HFT ends up undermining the role of public, transparent securities markets.

    1. This comment has been removed by the author.

    2. I don't know why you think HFT has anything to do with that. Front running was an institutional problem long before HFT existed.

      Dark pools allow large blocks to execute without the order size being known. When someone knows how much you're bidding/offering, they can try to game that to your disadvantage. The first attempts at solving this problem were done by regular exchanges when they implemented what are known as "iceberg" orders, in which you're allowed to specify some amount less than what you're actually bidding (never more) that will show up in the order book. But icebergs were eventually games as well as players would send small, incremental orders trying to test the actual size. (this order type still exists however)

      Please provide some evidence to support your hypothesis. I don't think you have anything to support it.

    3. Yes to some extent sniffing out big orders in advance was possible before HFT, but HFT automates and improves the process. HFT also uses past big orders/moves as signals to buy comparable assets. If you're trying to argue that the move towards dark pools has nothing to do with avoiding HFT, you're being silly.

    4. Execution on dark pools isn't point and click. Dark pools are a feature of HFT, not a bug.

    5. > but HFT automates and improves the process

      Tom, front-running was incredibly prevalent in the day's of NYSE specialists. Or how about the futures pits? How do you think all of those guys with high-school educations made so much money for themselves.

  7. Anonymous5:16 AM

    I don't think people talk about HFT in context of commodity trading. In case of electricity there really should be a fundamental value to milliseconds trading decisions as electricity cannot be stored easily and redirecting electrity may make allocation more efficient.
    I think people are biased against the strange unknown of millisecond markets.

  8. One highly overlooked impact of HFT is the market distortion caused by excessive layering and cancellations (this is exacerbated by the maker-taker/rebate model in the stock market). This causes artificially thick bid-ask levels which are to some degree illusory - particularly during moments of high volatility. This actually causes a decrease in volatility most of the time, followed by abrupt "Minsky moments" with magnified spikes in prices. Penalizing excessive quote/trade ratios and banning the maker-taker model would be a good way of keeping this in check IMO.

  9. Anonymous9:27 AM

    This is just 'jumping the que,' isn't it? What value is created by jumping the que? This activity used to be illegal, equivalent to trading on insider information. Tax it until it dies. And good riddance.

  10. Hopefully someone can help me out here, I've always viewed arguments in favor of HFT as muddy-headed at best, disingenuous at worst. There are two positives posited - increased liquidity and decreased transaction costs.

    On liquidity, firstly note that, by definition, these trades only last a fraction of a second. They are inserted between the seller and the long term buyer. Now, if long term buyers go away, so do the HFTs. That doesn't sound like increased liquidity to me. Another way to look at it would be to look at the total capital at risk in HFT operations relative to the market cap of the stock market (well not quite that but something of a similar order of magnitude). It's miniscule.

    On decreased transaction costs, again, I can't see how anything but the opposite could result from HFT. These guys aren't doing it for negative profit. Compare two scenarios, a trade between a seller and a long term buyer with and without a HFT sandwiched in between. In the first scenario, it took one transaction, in the second it took two. And what's worse, in the second the HFT skimmed some profit. So if we assume no variable cost to extra transactions, then the fixed cost has been spread over 2 transactions in the second scenario, making the immediate transaction cost half as much, but if viewed as a complete whole, the cost is the same, but buyer and seller have been made poorer by the amount of profit HFT was able to skim. And of course it's worse than that, because there is variable cost to transactions.

    What am I missing?

    1. That HFT is largely taking this money from human short term traders/market makers who used to do the same thing (albeit in timeframes observable to the human eye). In most cases I have no issue with this (although as a short-term human trader it costs me money). Take stat arb for example - is it really necessary to have humans maintain a correlation between S&P and Nasdaq tick-for-tick all day long? The problem is when a) individuals can pay money for a significant execution advantage and b) the times when computers are less efficient at pricing.

    2. Fair point - though I do believe that this difference in degree is large enough to have become a difference in kind.

      And the difference doesn't, to my mind, argue against my favorite proposed "remedy", namely a transaction tax. Given that we all currently effectively pay a toll to the guy that parked a big supercomputer next to the exchange, why not just pay that toll to a government that we all have a stake in.

    3. Andy,
      The commonly parroted phrase of "these trades last a fraction of a second" isn't really true. It doesn't make sense when you actually think about the mechanics of trading. And it makes it seem like people with fast computers are printing all sorts of free money — also not true.

      A person taking liquidity has to buy at the best offer. This means that unless that best offer is knocked out, the person taking liquidity is already losing on the trade. Further, even if the market does improve in the taker's favor, they can't immediately sell because what price are they going to sell at? The new best bid is likely the price they bought at. This implies that they're holding inventory for a while. And often they can't get out of their position when the price goes back against them.

    4. Andy, a transaction tax (at the rates commonly proposed) is a draconian response to this problem. It decimates liquidity. While this is not the only factor to consider, it still provides value (even to individual long-term investors).

      Also, illiquid products trade much more erratically and are much more prone to manipulation by large players. If the objective is to maximize the informational efficiency of markets, a fee for excessive quote/trade ratios will work far better.

  11. Anonymous1:04 PM

    Noah, this guy has his causality all mixed up. When a particular stock is more volatile, it becomes less liquid, while simultaneously more profitable to attack ( take liquidity ) - that's why multiple HFTs switch to attacking it rather than providing liquidity. Nothing to see here, moving on.

  12. It does seem as though complex order types may be a problem:


  13. Something I've been thinking about:
    From the 30,000 foot perspective I'm having a hard time seeing how a crypto-currency like BitCoin is any different than high frequency trading. Both require large scale computing hammering away at algorithms to generate perceptible value. They're both opaque to the laymen in showing how they generate utility and value. And if your point on HFT reducing liquidity is validated then they both have liquidity issues; e.g finding a BitCoin exchange as to waiting for an HFT to unwind.
    I suppose HFT are trading "real" underlying assets (although if they were working on synthetic products, that too would be debatable right?)
    While I'm sure anyone working on HFT would sneer at the thought of generating BitCoins (I'm sure a BitCoin miner would think the same at going into HFT). Aren't their net results the same?

  14. Noah, very interesting and timely post! You might be interested to know that together with my co-authors Sandrine Jacob Leal, Mauro Napoletano and Andrea Roventini, I have just finished a paper where we theoretically investigate whether high-frequency trading exacerbates market volatility and generates flash crashes (http://arxiv.org/abs/1402.2046).

    Our main result is that the presence of high-frequency trading increases market volatility and plays a fundamental role in the generation of flash crashes. The emergence of flash crashes is explained by the ability of high-frequency traders to siphon liquidity off and to massively sell. We also find that higher rates of order cancellation by high-frequency traders increase the incidence of flash crashes but reduce their duration. We infer that order cancellation strategies of HF traders cast more complex effects than thought so far, and that regulatory policies aimed at curbing such practices (e.g., the imposition of cancellation fees) should take such dynamics into account. I hope you might find this article interesting!

    1. Giorgio,
      I am a huge proponent of ABMs and I appreciate your efforts in modeling. However, after a cursory examination of your paper I have some concerns.

      First, I think your paraphrased citation of the Kirilenko 2011 paper might be misrepresenting its findings.

      Second, you do not define "flash crash" nor present empirical findings (yours or otherwise) of their pervasiveness.

      Third, the papers you're leaning on do not differentiate between what is known as a "fat finger" trade and a flash crash. Consistent with the Kirilenko paper, HFT is not the cause of the flash crashes, it merely exacerbates volatility when the market is already sweeping. (The Sornette 2011 is particularly weak because of this)

      I think your modeling efforts would be best directed at investigating how HFTs respond to these fat finger trades.

    2. Dear ZHD

      Thanks very much for your comments.

      We show that flash crashes are caused i) by the synchronization of HF traders on the sell side in a situation where ii) bid-ask spreads are large. These two elements are in line with the description of the flash crashes contained in Kirilenko et al (2011) paper, although you are right, the empirical paper of Kirilenko only says that HFTs were only exacerbating downward price movements.

      I disagree with you that we do not provide a definition of flash crashes. We define flash crash at p.5 "as drops in the asset price of at least 5% followed by a sudden recovery of 30 minutes at maximum". This was based on the evidence about the size and duration of the May 6, 2010 flash crash reported in Kirilenko et al. (2011) paper.

      Regarding flash crash pervasiveness, so far this aspect has been left for future research since in the present model we focus only on one market. This aspect certainly deserves more attention.

      Finally, I appreciate your point about fat finger trade and how HF traders could respond it. However, our point in this paper was different. It was more about investigating whether there were features of HFT that could generate flash crashes. We show that some directional strategies employed by HF traders may lead to large bid-ask spreads and therefore set the conditions for the occurrence of flash crashes. It's also because of this important element that we claim, indeed, that HFT can cause flash crashes.

    3. Thanks Giorgio,

      My apologies for missing your definition on the first read. That's usually a good signal for me to do a closer inspection.

      The point I was trying to make about the fat-finger vs true flash crash is that without knowing which is which, your model is unlikely to reflect reality.

      From what I see, it looks like you have a good start to the idea. But I don't think the market microstructure is correctly specified to reflect reality.

      Do you have market makers? Without modeling this you're of course going to get crashes.

      Further, allowing HF traders to increase their size based on available book liquidity is of course going to lead to crashes as well. Setting a hard position limit is insufficient because that's simply not how firms allocate capital.

      Given that you allow long and short positions, you should also be seeing flash rallies as well.

      There is a lot more that doesn't accurately reflect market microstructure. But you're on your way. The great thing about agent models is their ability to be improved.