In a first for Flirting with Models, my guest this episode is anonymous, going only by the handle LightSpringFox on Twitter.
Mr. Fox is a quantitative trader who works in crypto market making at MGNR. Mr. Fox did not begin his career in crypto, nor even in market making. Rather, his background is in traditional equity factor investing, and so we spend a good deal of comparing and contrasting the low- and high-frequency domains. We also discuss the nature of market making edges, the unique risks of high frequency, how crypto and traditional finance market making deviate, and what Mr. Fox considers the “hardest problem in HFT.”
Without further ado, please enjoy my conversation with LightSpringFox.
Corey Hoffstein 00:00
Okay, are you ready? Yep. All right 321 Let’s do it. Hello and welcome, everyone. I’m Corey Hoffstein. And this is flirting with models, the podcast that pulls back the curtain to discover the human factor behind the quantitative strategy.
Corey Hoffstein Is the co founder and chief investment officer of new found research due to industry regulations, he will not discuss any of new found researches funds on this podcast all opinions expressed by podcast participants are solely their own opinion and do not reflect the opinion of newfound research. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of newfound research may maintain positions and securities discussed in this podcast for more information is it think newfound.com.
Corey Hoffstein 00:53
If you enjoy this podcast, we’d greatly appreciate it. If you could leave us a rating or review on your favorite podcast platform and check out our sponsor this season. It’s well it’s me. People ask me all the time, Cory, what do you actually do? Well, back in 2008, I co founded newfound research. We’re a quantitative investment and research firm dedicated to helping investors proactively navigate the risks of investing through more holistic diversification. Whether through the funds we manage the Exchange Traded products we power, or the total portfolio solutions we construct like the structural Alpha model portfolio series, we offer a variety of solutions to financial advisors and institutions. Check us out at www dot Tink newfound.com. And now on with the show. In a first for flirting with models, my guest this episode is anonymous, going only by the handle light spring Fox on Twitter. Mr. Fox is a quantitative trader who works in crypto market making it mdnr. But Mr. Fox did not begin his career in crypto, nor even in market making. Rather his background is in traditional equity factor investing. And so we spend a good deal of time comparing and contrasting the low and high frequency domains. We also discuss the nature of market making edges, the unique risks of high frequency, how crypto and traditional finance market making deviate and what Mr. Fox considers the quote, hardest problem in HFT. Without further ado, please enjoy my conversation with light spring Fox. Well, this is a first for me. At truly anonymous guest light spring Fox. Can I Can I call you Mr. Fox? Yeah, that’s perfect. Mr. Fox, welcome to the show. Welcome to flirting with models. This is going to be a fun episode. Let’s just start with the obvious question. Where did the name late spring Fox come from?
Well, first of all, thanks, Cory. I appreciate you having me on. It’s great to be here. So the light spring Fox name a couple of years back. For your listeners who have been in crypto for a while bitmax was kind of the biggest derivatives venue. And they had this public leaderboard where you could either choose your handle, or it would be like randomly generated three words. And at the time, the top of the leaderboard, was this randomly generated three word combination, heavy autumn Wolf. And I forgot what their p&l was, but it was like, must have been in the 10s of millions in dollars. And at the time, I was just thinking, wow, this is like a real professional. There’s real money to be made here. And I want to approach a level that’s somewhere near this, but like, I might not be the heavy autumn Wolf, but I’m happy to beat a light spring Fox.
Corey Hoffstein 03:55
I love it. I love it. But you weren’t always in crypto. So before we dive into the world of crypto and sort of the quantitative strategies that you do work on, let’s actually start with your background. You’re actually come from a more traditional quantitative finance world.
Yeah, I would say I had a pretty traditional start. I studied engineering and I graduated in 2018. So I’m still not super late career and out of school I worked at a quantitative equities manager for two years. Picture like in AQR, but it wasn’t AQR but similar strategies. So long horizon, infrequent rebalances but still very quantitative and automated in their approach. So was writing a lot of code and doing a lot of data analysis there.
Corey Hoffstein 04:49
Now, you are currently employed at a quantitative investment firm in the crypto space I have to ask why continued to remain anonymous.
Well, you know, to me Krypto, just based on what I’ve seen in my time in the space, it seems like a super optimal thing to steal. And that’s an unfortunate kind of truth of the state of the current industry where if someone gets into your bank or charges your credit card, a lot of the times that somehow ensured it can somehow be reversed. You call the right people and you get your money back here. If someone gets into your crypto wallet, there’s very little shot of getting that money back. And there’s been some high profile targeted hacking attacks. And to me, it’s like, it’s just an expected value problem. Like I’m not I’m not the super paranoid guy, where no one knows who I am like, if you DM me on Twitter, and I trust you, I’ll meet up with you and you’ll know who I am. But just the expected value of having my real name out there versus the probability of a specific targeted attack occurring, because someone knows that I handle large quantities of crypto doesn’t seem like an unreasonable risk to be taking to me,
Corey Hoffstein 06:07
do you think that as the crypto space evolves, operations, security will evolve in such a way that that will no longer be necessary? Like that sort of positive Evie decision for you to remain anonymous may not be necessary anymore? Do you think it’s in the nature of crypto itself? That that’s sort of always going to be necessary, potentially,
you know, I would hope not like from a mass adoption standpoint, it’s not really a great state of affairs that you’re afraid to have your name out there. To be clear, I think you can have the right operational security and the right setup where your risk is very low. I mean, there’s a number of public figures out there who are in the crypto space very publicly with their real name out there. And they have the security practices. And they put a lot of time into those things, where they’re not worried. But for me, because I’m like interacting with order books, and not really doing venture deals and stuff. I just don’t want to bother with having to maintain super high levels of operational security all the time.
Corey Hoffstein 07:14
So let’s dive into the nitty gritty a little bit. Because your career transition hasn’t just been from traditional finance to crypto, it’s also been from these more long horizon sort of equity factor strategies to higher frequency market making type strategies. Thought it’d be really interesting if you could compare and contrast maybe the difference between that short horizon and long horizon quantitative trading?
Yeah, absolutely. I guess a good way to frame this is as like a set of reasons why higher frequency horizons are better, and reasons why they’re worse. So I think let’s start with the good in high frequency data is super abundant, our systems will generate more data in a day than easily then the entire, like, fundamental balance sheet history of all US equities ever. And from that standpoint, it’s very nice modeling environment where you don’t have to worry about overfitting so much changes are statistically significant in production very fast. So if you make a change, it’s like instant gratification, you don’t have to wait months or years to be certain that it made a difference, the variance is just a lower. And another pro that comes with that is the live equity curve you end up with is very impressive. If you do this, well. If I showed, like our live equity curve to someone from my old job, they would say what mistake did you make on this backtest that you’re buying the stocks that went up, it would just look like look ahead bias. Now, the cons that come with that is there’s very like vicious adverse selection in these trades. And everyone wants that live equity curve. And there’s only a limited capacity for the number of those juiciest trades that you can take. So you’re competing with smart teams all around the world who are essentially going for the same opportunities. And as a result, the alpha decay is much faster than that longer horizons. And the capacities tend to run lower. You can’t just abstract out the market and work on some statistical research for long periods of time, which I may have done in my previous role. You really have to be thinking about execution every step of the way, because it’s so much of where the edge stands here.
Corey Hoffstein 09:42
Well, let’s talk about edges for a second. So Michael Mobizen, sort of famously classified edges into four different categories. There’s the behavioral edge where you’re taking advantage of the misbehavior of others. There’s the analytical edge where you’re, you have the same information as everyone else, but you’re able to analyze it in a unique way to dig out, maybe a unique insight. There’s an informational edge that you literally know different things than other participants. And then there’s a technical edge, which might be something like you can execute more efficiently or with less impact or more quickly than your competitors. I’m curious where you think the edge really lies in crypto market making?
Right? So I think that edges are largely to operate within those four categories. I think that edges are largely analytical and technical. Because especially at the high frequency horizons, where microstructure is really what matters, the problem kind of collapses into fewer sources of variance. And those sources of variance aren’t a secret, you have public order books and a bunch of correlated instruments, you have the order flow going through those. So there’s really, at the microstructure horizon, you’re not really finding an informational edge where there’s something that you know, moves the market in the next one second, that no one else knows, well, you’re more focused on taking all the existing information, and analyzing it better. To go with that there’s a strong technical component where you need to execute on this very fast because to some extent, your model can be better and more accurate. But you’re still going to be correlated with all the other participants playing the same game. And that’s where that execution edge comes in, where you just need to be very, very fast, whether that’s communicating information between different places where you’re processing time within one place. And as for the behavioral component, I think, maybe as the market has matured, this has disappeared, at least as a sole source of profitability. Because maybe a year or two ago, you could, for example, just quote really wide and in liquid instruments, and every once in a while, some maniac would come in and just pay like, we call it a PE through so they just blast like $100 million into that book, and you’d catch 100 pips or whatever. But now increasingly, like the markets have gotten more efficient, liquidity has gotten deeper, and that sort of bad behavior. Maybe it happens once in a blue moon. But you can’t really rely on it as your sole source of edge.
Corey Hoffstein 12:36
I want to push into that technical side a little bit. Because I do know that in traditional finance, right, there’s a big arms race in high frequency trading around colocation, and getting down to the metal versus in crypto, a lot of the centralized exchanges are hosted in AWS, Amazon cloud servers around the world. And the arms race seems to be a little bit different. There’s some AWS networking knowledge that can help you maybe get close. But it does seem like there’s limits to the colocation ability, and ultimately just sort of that in the network speed, that’s achievable. And then for folks who maybe do trade, on the decentralized side, right, you’re ultimately limited by block speed. So I’m curious as to when you think about that trade off between analytical and technical are there limiting factors around the way centralized exchanges are designed in crypto, that prohibit really the same emphasis on the technical as you tend to see in traditional finance.
So I would say the emphasis on the technical is equally strong, but it’s a different flavor, it’s a slightly more creative, and I would almost call it more of an art than a science. So in traditional finance, it’s a very surgical discipline and a very constrained optimization. As far as HFT goes, because you have the situation where everyone has the same length cable to the matching engine, the matching engine is very deterministic, everyone has heard those stories. And what ends up happening is you’re implementing in hardware, your entire trading strategy and just counting nanoseconds. And the task is like very defined. It’s like every nanosecond, we gain here is a now second, that puts us into first place and it doesn’t matter how much you win by well in crypto, there’s a lot of what we call jitter is the jargon, which basically means there’s a lack of determinism and how long things take the process. So even if you’re, let’s say slightly ahead, you might end up behind by the time all is said and done and the matching engine processes and the orders hit the book So understanding those unique behaviors, understanding the quirks of each exchanges API, understanding how to get information from one AWS region to another fast is more the game here. And a lot of the systems weren’t built for like this use case. So in traditional finance, when you’re co located in a data center with your FPGA, or whatever, every component of that pipeline has been purpose built. For the purpose of submitting orders to a matching engine fast, while AWS is built to be good for most things, it’s good for hosting Instagram, it’s good for doing big machine learning tasks in the cloud. But no one was really thinking about this, when they built these exchanges. It’s just the easiest way to scale something fast on AWS, especially if you don’t know how big it’s gonna get. And you don’t want to invest in tons of bespoke hardware upfront. And all of those factors summarize to a large technical emphasis, but a very different flavor of kind of creative optimization, working around a system that wasn’t necessarily built for the purpose of highly competitive trading.
Corey Hoffstein 16:17
One of the things you skipped over, in discussing your background is that you actually struck out on your own for a bit, doing some independent market making in crypto between your sort of career in traditional finance and working at the firm you’re working at now. I’m curious, in those early days, what was your most educational experience,
I think the most educational thing you can do is to just put something out into the market and see what happens. And for us that was like we had a simulator. And we just very quickly realized that every single gap between the simulator or model and production would just realize in maximally adverse fashion. And by that I mean, when you fit a model, you expect the error term to be centered with mean zero. And like sometimes you’ll make more money than you expected. Sometimes you make less. In this case, any source of uncertainty was like maximally negative. And that’s like a very hard and frustrating lesson. But it really taught us about the game we were playing when we were independent. And the other thing I would say is, there’s kind of a almost like a gospel, passed down through generations in this industry of various heuristics and tricks, and just things that are learned over the years that are very hard to find publicly. And I was lucky enough to kind of maintain a friendship with someone who used to do this very competitively, and has since moved on to other ventures. And as a result, they were a little more comfortable sharing the nature of the kinds of things they did, and I would just have very long email chains with this person, where I would just share like, okay, so we have this exact problem, here’s the data behind it. And they’d say, Oh, of course, that’s like, very standard stuff. But for us, we’re just learning, usually the hard way. So that kind of like really sped up the curve for us. So yeah, I owe a lot to that person. If if you’re listening, you know who you are.
Corey Hoffstein 18:32
If I were to shadow you for a day, what does the day of crypto market maker actually look like? Broadly,
I would categorize the job as distracted coding. I’m always working on improvements to the system. I always have code in front of me whether that’s like Jupyter notebook or just more production infrastructure code. And then I’m also kind of monitoring the systems and what they’re doing, we have a number of UIs that give us transparency as to what’s going on. And most of the time, everything will run in a manner where you don’t really have to be watching it like a hawk. But every once in a while an interesting event will come up and you’ll zoom in on, think about what happened. I guess to give a more specific example, let’s say the market moves 100 pips in like one second, which happens frequently in crypto, right. And you would just kind of look at that and see if you made all the trades that you wanted to see if someone was faster in certain cases, and then maybe follow up with a more broad data analysis because any single event is more of an anecdote. You don’t want to introduce changes based on that. So then you would look at the historical telemetry, the historical exchange data and see what’s been going on in the past. And once you affirm that there’s a pattern and like maybe a meaningful improvement you can make, you would then start to introduce changes to production with a new feature and a new element of the model, something to make us faster.
Corey Hoffstein 20:11
How frequently does that event analysis lead to a change in production?
That’s a good question, I think you’d usually want to observe something a few times. So maybe like 30% of all events constitute some change that you want to make, as opposed to just the market moved, and everything kind of happened as expected. And then from there, you’d probably want to observe that same issue a few times. But it’s a lot really, because you can always do better, even if you made money in a situation, you could have always made more, you could have always been faster, you could have always been more accurate. And your competition is doing this the whole time. Also. So if you just let the system sit stagnant, and don’t make these improvements based on the observed trading data, what’s going to happen is over the course of months, your alpha is just going to decay to zero, and then it’s going to decay to negative because everyone else is doing the same thing. Who’s playing this game?
Corey Hoffstein 21:09
So without leaking any alpha? Can you talk a little bit about what these changes might look like? Like, is this just parameter tuning? Or is this building whole new algorithms? Like what do you what are you actually changing after you do this event analysis?
Yeah, so I wouldn’t call it so much as tweaking parameters, that’s something you would find more unlike an options desk may be where you’re constantly nudging the vault surface and trying to inform your system as to like your view on the market, were more just kind of finding places where the system doesn’t react in a way that we would expect to based on like our economic intuition. And sometimes that might be as simple as even bad data coming in. So let’s say some venue publishes an over the counter block trade for massive size. But the data is stale, it’s time stamped 15 minutes after the actual transaction occurs, because that’s how they choose to catalog this particular type of block trade. So what’s going to happen is, if your system has no way of identifying that this data is stale, it’s going to pick up all this order flow that it thinks is happening right now. And consider a Bitcoin to be where five MIPS more whatever the case may be, and do a bunch of trades based on that. So in an event analysis, you would go back and look at this source of input, you would isolate it as the culprit for what really moved things for you, and figure out a better way to handle that piece of data or maybe ignore the stale data. Or it could be the opposite, where you see a trade that based on your economic intuition, should have moved the market, or you see a change in the order books that should have moved in the market. And you see that this was the first sign of change in the market. And you’re wondering why your model isn’t picking it up. Or maybe it’s picking it up, but it’s not weighing it heavily enough. So the model is still very automated in its approach. But you might be nudging how to value certain inputs or trying to figure out why certain inputs were valued the way they were. And a lot of this just comes down to being faster also, to clarify, like when the market moves, you’re, it really comes down to changing your quotes that are the wrong price, and removing other people’s quotes that are the wrong price. So if, for example, you tried to change your quotes, but you weren’t fast enough, a question you might ask based on that event study is what is the competition doing to be faster? Are there certain optimizations they have in place? What do we think those optimizations might look like? And how can we obtain them ourselves or obtain something better?
Corey Hoffstein 24:10
You said that you don’t necessarily do parameter tuning like someone who does a high frequency options market making desk that might nudge parameters based on their view of the ball surface, but you did sort of seem to say that you might nudge other parameters. I’m curious about this idea of of parameter tuning, because it is something that comes up a lot when you talk to people in the market making space. Can you maybe elaborate a little bit on what parameter tuning is and whether you guys do it or not, and maybe why it’s done. If it is,
right. As I stated earlier, I think our system leans quite automated. We could leave it alone for days at a time and it would be just fine. But the reason changing parameters is such a popular thing in high frequency is, again, this sample size thing I alluded to, like sample size accumulates really fast. If your long short equity portfolio has a good week, you’re not gonna, like, go out and borrow money to put into that strategy. That’s variance, right? While in in high frequency, that might actually be a reasonable thing to suggest. So a lot of the parameter tuning we do isn’t valuation level or Lincoln options. As you mentioned, it’s more like, Okay, this strategy is doing well, we have a good edge on the market, we need to push that advantage while we have it. So we might start to size up quite aggressively, we might start to command lower edge, just because we think our idea of fair value is so accurate, that we’re willing to have a tighter confidence interval. So a lot of parameter tuning. Even in the most automated system, you’re still going to have those parameters like how big do I want to trade. And then I guess to kind of hit the opposite side of that equation. If there’s a scenario where you don’t think you have a lot of edge, you might dial things back. So for example, if there’s a big news event coming out, and you’re not sure how your system is going to respond to that, you might pull some of your orders, dial back the overall risk tolerance and just let that 32nd news release happen, and then get back into the market.
Corey Hoffstein 26:36
We did a pre call preparing for this interview, one of the things you said to me offhanded was that the hardest problem in high frequency trading is, quote, conditioned on getting filled, do you still have an edge? That That question was sort of at the core of everything you look to do? What do you mean by that?
Right? I think a good way to frame this is to introduce this term of model edge in like a slower factor equity strategy. If you have high R squared on stock returns, you’re like 80% of the way there, right, you just need to put this in a portfolio optimizer long the high return expectations short, the bad return expectations, call your prime broker and rebalance those positions every week, month, or quarter. And hopefully, they’re liquid enough that your market impact isn’t too big. Well, in high frequency model edge can have almost zero bearing on your p&l. And I’ll motivate this with an example. So there’s this open secret of book pressure, right? So you just formulate an input based on the quantity on the best bid and the quantity on the best ask. And if you were to do a regression of this book imbalance on the direction of the next tick, in most instruments, you would get an r squared that inequity factor Quan would think is incredible. But the problem is conditional on getting filled, you actually have no edge in that situation, because what’s going to happen is, this signal alone won’t be enough to cross the spread to pay both the spread and the fee differential of removing liquidity. But if you try to trade the signal passively, you’re gonna wait for this imbalance to build up, because by definition, that’s how your signal works, you’re gonna add your order on the best bid, in the case of a positive imbalance. And you’re going to be at the back of the queue on the best bid. So the only time you’re gonna get filled, is that subset of times that you’re wrong about the next tick, because everyone on the best bid in front of you has to get filled first. So that’s like a motivating example of how you could have a very high R squared, but you’re actually conditional on getting filled is definitely negative with this
Corey Hoffstein 29:11
in the low latency space when systems go down, and it’s usually just a headache. You talked about quarterly rebalanced equity factor portfolios. If your systems go down, you normally have hours if not days, and weeks, in the worst case to get them back up and running and get that portfolio rebuilt. In the high frequency space, those system errors can lead to situations like we saw a couple of years ago where Knight Capital Group burned through over $400 million in just 45 minutes for when they pushed some code errors. Actually, I think it was technically their server failed to update was the problem and yet three servers running on new code and one server running on old code. Anyway, neither here nor there. The the real question I’m getting at here is how do you think about man During technology risk,
it’s a few different things. I think one with KCG with my capital, the scenario you describe, if you go and read the SEC breakdown of that event is, the engineers over there, we’re trying to figure out what was wrong for 45 minutes, maybe I don’t fully understand the situation. So I don’t want to talk too badly about it. But really, our philosophy all the time is, if anything looks off, just shut it down immediately. So turn it off first, ask questions later. And some of the time that it will be automated with like sanity checks. So you might even give up some subset of like really bizarre opportunities, because they trip a sanity check, you’re gonna say, there’s too high of a probability that my market data here is wrong, that I’m not even going to, I’m just going to turn off Cancel all my orders and wait for a human being to check on it. The other end of that is if we always have someone watching from the corner of their eye. So if there’s something that gets passed all of our sanity checks, and you’re almost conditioned, based on the UI we have built out too, as to what’s normal, and what’s not. Because you have that visual feedback all day, you just shut it down right away, and then figure out what’s going on after. Obviously, hopefully, things never get to either this point, either the automated turning off or the manual turning off. So the other piece of that is a really rigorous code review process, where at times, it can even feel like TDs people will be nitpicky over certain points. But in reality, it’s all a really good thing. It makes everyone a better programmer, it makes sure that all the code that gets into production is readable. It’s done to a high standard of quality. So you have sign off from several members of the team, basically, whenever something goes into production, and I think a lot of productive discussions come out of that as to making the system safer.
Corey Hoffstein 32:07
Another unique compare and contrast when you look at the traditional finance space versus the crypto space is that in traditional finance, having a security listed on multiple exchanges tends to be more of an exception than the rule. Whereas in crypto, it tends to be more the rule than the exception. You tend to find Bitcoin, for example, trading on every major centralized exchange and decentralized exchange. How do you think this presents both opportunities, as well as challenges for market makers in the crypto space,
I think the fragmentation presents obvious opportunity in the sense that if you can account for all the unique differences between the different venues between the different instruments that are essentially exposure to the same thing, there’s edge to be had there, right, because it’s like a hard distributed systems problem to solve. We spend a lot of time and engineering on keeping it running smoothly, and getting information from one place to another. So if you solve the problem, right, there’s opportunity because it’s a hard problem to solve. I guess the challenges are in the sense that these things aren’t fungible, they’re actually different instruments. So even for something like perpetual swap, the most popular kind of liquid future, what happens is exchanges will have slightly different funding rates, they’ll have slightly different funding intervals, the contract will have a different index that the swap is priced on. So you need to account for all of these differences in a generalized manner. Definitely not by hand, because there’s like hundreds and hundreds of these things. So building a system that does that, well, is challenging, and you’re always kind of hunting down the edge cases of like, really did this exchange decide to do it this way. And we had something that worked everything everywhere else. And there’s complete lack of standardization where this one venue decided to do it a different way. But there’s too much edge there for us not to trade. So we’re gonna have to read generalize, to cover that edge case. So I think that’s where a lot of the challenge comes from.
Corey Hoffstein 34:25
We’re starting to see a lot more high frequency trading firms enter the crypto arena, what domain knowledge do you think is portable from traditional finance? And what do you think will end up surprising these firms as they sort of make their foray? Right,
so I think it kind of happened in in an interesting way and were over the past year like volumes really skyrocketed and everyone started building these things out, and probably going live into the The later portion of 2021 or early 2022. And now we’re in an environment where volumes have kind of gone down, actually. And edge has gone down and the market has gotten more competitive. I think first of all, people will just be surprised of how hard of a game they will set out to play at the start of 2021, which is, by the time the last lawyer signs the last paper, and the last approvals come through. It’s a very different market, because now every single major proprietary trading firm is at least looking at this or probably doing something. So I think that’s one of the biggest surprises, I guess the other surprise would go along with what I mentioned earlier on the technical edge, where the less constrained optimization, were being fast as important, but you optimize for it in slightly different ways. And if you’re used to trading on like, super deterministic CME, it might take a little getting used to. But aside from that, I think a lot of Trad five quants will find themselves at home here, just because an order book is an order book. To be honest, I think crypto native prop firm is kind of a misnomer. You look inside a lot of crypto native firms, and it’s mostly people who are like ex traditional HFT. So as far as that goes, I think the markets themselves are similar. Can you
Corey Hoffstein 36:31
expand a bit on how the landscape has evolved in centralized exchanges in crypto, and maybe how that’s impacted the market makers over the last couple of years?
Sure. So I think first of all, the exchanges have gotten better. So there’s still a far cry from the super high performing Trad fi exchanges. But the exchanges have gotten better. And it’s allowed for more complex systems to be in place, you’re not waiting five seconds for your order to confirm. So in that sense, I think the market is always getting faster, and it will continue to get faster as the exchanges get better. And the competition increases. I think the obvious arbitrage is are kind of harder to come by. Everyone talks about how Alameda got their start, I believe, with this Japan trade, or was it Korea? Either way, there’s just like a 10%. price difference, right. And I think that it would have been great to be doing this at that time. But I think those opportunities are very hard to come by now with the attention that’s come to this market. But, you know, in a sense, I think there may also be some consolidation and exchanges, especially if we continue to go through this period of lower volumes. Some of the smaller exchanges might roll up into the bigger ones. But that’s just speculation on my part. I don’t know if a lot of deals like that have gone down yet.
Corey Hoffstein 38:04
I frequency trading has historically been a surprisingly polarizing topic in traditional finance. What do you think the greatest misperceptions are about high frequency market making?
I think the biggest misperception at least among quantitatively oriented people, since that’s who your audience is, is probably that these strategies consist of execution edge, unlike really dumb trades. Atomic arbitrage is where the price is different between two exchanges. Or just like some get in front of someone’s order, where like, sure that segment of the market kind of exists. But in any liquid competitive market, you actually need to get far more complex and nuanced. And than that, again, with the abundance of data, you can have very complex models that live in rich parameter spaces. The things we trade here are much higher dimensional models than what I traded in equity is even though equities actually have far more sources of variance. I guess that’s a long winded way of saying that we’re not just being faster to atomic arbitrage is here are using some structural advantage to do a really obvious trade in atomic arbitrage just kind of disappears as long as someone is willing to do it non atomically. If they’re not going across the spread to hedge somewhere else there can give a better price and warehouse the risk and everyone who does well in this market is willing to warehouse the risk. And so I think in that sense, a lot of people would be surprised to actually step into a high frequency trading firm and see the kind have edges were chasing?
Corey Hoffstein 40:01
One of my favorite questions to ask of investors or traders of any kind is to ask them to sit across the table and take an allocators perspective. Now, I know most allocators never get the opportunity to invest in a crypto market making firm. But let’s pretend for a moment that you had that opportunity that you could invest in a crypto market making firm. What due diligence questions would you ask?
Sure. So I alluded to this a little bit earlier, I think the main things should center around sustainability and scalability. So sustainability, meaning alpha decay, and scalability. Meaning, how much capital can you actually trade with these strategies? For example, if you have a firm that derives most of its p&l from the most liquid instruments on the most liquid venues, you know, that they have to have been continuously improving. And they’re probably trading more high capacity strategies just because the bulk of the volume is there. So because of their track record, in the most competitive markets, they’ve kind of proven their ability to continuously beating the alpha decay, even though it’s an inevitability. And because of their where the volume is, they can probably handle more money. On the other side of that there’s firms that derive most of their p&l from more esoteric trades. So maybe they’re trading tail end all coins on backwater exchanges and making a lot of money there. And that can be interesting in its own way. But you have to ask yourself the question, if there’s a sustained bear market, where these backwater venues roll up into larger ones, or these tail end, all coins just all go to zero. What is this team going to do? And is there even any capacity they can retain? But I don’t think it’s fair to like totally right off that style. Because if someone has a track record of always finding the next esoteric opportunities, they haven’t just been trading the same one for six months. There’s also firms that trade in that style that I would definitely invest in just because even though they’re not trading the most competitive markets, they have a knack for always finding a place to trade. So even if the all coins they currently trade all go to zero, they’ll find something interesting to do somewhere,
Corey Hoffstein 42:37
you made a pretty significant career change from lower frequency, equity factors to high frequency crypto market making, what advice would you give to someone who was either looking to make a similar jump or just start a career in high frequency market making?
My advice would be that, unfortunately, it’s pretty hard. When I was in lower frequency, I went in interviewed, and the overall sentiment was, what you do now is pretty unrelated. So the best way to come in is, when you graduate, the way the recruiting is done in these places is you come in as an undergrad, and they teach you what you need to know. And those are like battle tested training programs that are pretty good and what they offer, because most of these firms just want to train you up themselves. They don’t want to hire like someone who’s been doing something else unrelated, they might hire like an experienced trader, from a competitor that’s also interesting to them. But otherwise, what’s most interesting to them, aside from an experienced hire from a competitor is someone who’s a fresh graduate of undergrad or PhD. Now that being said, it’s not all doom and gloom, because I think crypto has really changed this landscape in terms of its accessibility. If you look at my own kind of HFT project that I did with one partner, we would have never even been able to consider independently trading something like equities or futures in the US, the costs, the colocation the data feeds, it would have just all been super prohibitive while in crypto, like it’s really democratic in terms of who can access the order book who can access the API. And sure there’s fee tiers, but I would even say, the next best way to learn aside from going to an hft firm out of undergrad is to just put $100 into crypto, trade minimum quantity on some altcoin where that’s like 10 cents or 50 cents that will get you Do a lot of trades even if you’re losing five pips each, right. And because it’s so accessible, you’re using the same API that the professionals are. And you might say, Oh, I don’t have the fee tiers, so I can never make money. And I say, lose your $100 and retro actively apply the highest feature to your life trades and see if you would have made money. And if you then go to a proprietary trading firm and crypto and say, I lost money, but if I applied the features that you probably had, I was making money, they’re going to seriously consider hiring you versus someone who came out of doing something unrelated. So I would say in that sense, because of the openness of the crypto markets, it’s now easier to break into HFT than it ever was previously.
Corey Hoffstein 45:50
The question I’m asking everyone, at the end of each episode this season, is what has been your luckiest break in your career thus far. And I know you’re still very early in your career. So there’s perhaps a little less to reflect upon. But I’m, I’m curious when you look back at sort of your biggest breaks things that have really opened up your career for you, if there’s anything that really stands out,
you know, I have to really look at Twitter. So I’ve never been a social media guy. I’m not on Instagram, I’m not on Facebook. And I can’t even say for sure why. But I saw that there was this community on Twitter. I made a Twitter account and just started posting my observations on the market, I could be a little more transparent when I was independent than I am now where I’m handling other people’s alphas. And it’s just crazy what came out of that. If you told me that making a Twitter account would result in the things that it did for me, I would have never believed you. It’s unbelievable who you could connect who you can connect with on this website. And I highly recommend posting things on it. Now, obviously, there’s a diminishing return curve. So you don’t spend all day on it. But yeah, I would say I would say if I look back at like all the little decisions, making the Twitter account, it seemed really lucky in a way how things went down with that.
Corey Hoffstein 47:23
Well, Mr. Fox, this has been a pleasure. I really appreciate you coming on and while anonymous willing to share and impart some wisdom for our listeners.
Thanks Cory. Great chatting with you.