In August 2021, Hummingbot's interview series with community members sat down with Eugene Tartakovsky, co-founder of 3Jane. The conversation traces how a software developer went from forex and Ethereum mining to systematic crypto trading. Along the way, it shows what old books, open-source tools and getting paid to trade had to do with it.
This page preserves the full video, a topic index and an edited transcript.
Watch the full interview
- Video: Interview with Eugene Tartakovsky | Community Spotlight
- Publisher: Hummingbot
- Published: August 19, 2021
- Duration: 45:21
Interview with Eugene Tartakovsky
The conversation runs from Eugene's first trades to the reasoning behind 3Jane's strategies. Eugene explains why brute-force optimization and price-predicting machine learning failed, why classical strategies worked better, and why finding the right market matters more than the strategy itself. The second half covers Hummingbot, custom infrastructure, liquidity mining campaigns and decentralized exchanges.
Topic index
Each topic opens the full video at the point where it starts:
- Origins, from 0:59: forex, Ethereum mining, ICOs, the 2018 crash and the books that followed.
- Strategy selection, from 13:32: brute-force optimization, machine learning and the move to classical strategies.
- Team and scale, from 24:58: where small players compete, 3Jane's team size and its monthly volume.
- Hummingbot, from 28:11: how 3Jane found it and which off-the-shelf strategies it used first.
- Custom infrastructure, from 30:08: custom pricing, speed, documentation, liquidity mining campaigns and decentralized exchanges.
Edited conversation
The transcript has been edited and condensed for length and clarity. Headings follow Hummingbot's published chapters, and each timestamp links to that point in the full video. Hummingbot marks the host's questions and comments; Eugene marks Eugene Tartakovsky's answers.
0:15 Introduction of 3Jane
Hummingbot: Welcome to another edition of our interview series with members of the community. Today we have Eugene from 3Jane. 3Jane is a Moscow-based quantitative trading firm that has been using Hummingbot and participating in Hummingbot Miner.
Hummingbot: They're a good example of what we're trying to do with Hummingbot: create software that lets anyone in the world do what historically only people on Wall Street could do. 3Jane is based in Moscow but trades crypto markets all over the world, using Hummingbot as well as software they've developed themselves.
Hummingbot: Eugene, can you tell us how you got into crypto and the origin story of 3Jane?
0:59 Backstory of 3Jane
From a software career to forex
Eugene: Sure. I'm a software developer, with a long career as a programmer and CTO. Around 2014, in the middle of my career, I got a bit tired of selling time for money. So I started reading and thinking about what else I could do besides a professional career, and I got into investing and trading a little.
Eugene: I started with forex. It wasn't an ideal result. I made money in the end, but later I realized how much risk I was taking. I was betting my whole capital, which could collapse to zero, to make maybe 20% a year. So it wasn't a good game.
Managed forex accounts
Hummingbot: Were you a quant at the time? Did you have your own algorithm?
Eugene: No, not back then. In Russia there's a thing called a PAMM account, where individual managers run forex portfolios that other people can invest in. What I learned is that the people running these accounts usually take unreasonable risks, even when the result looks profitable. I stopped while I was still in profit.
Looking for something better
Eugene: I kept reading. By about 2016, I had a better sense of what one can invest in, and what risks and profits are involved. I didn't much like the opportunities that are generally available, and Russians can't access most foreign brokerage accounts anyway:
- The stock market is a very specific game with huge drawdowns. You either pick winners or sit on about 5% a year, aside from 2021, apparently.
- The bond market is even more boring.
- Real estate requires a kind of expertise I don't have.
- Commodities aren't very popular outside the US, and they're not very liquid.
A gaming PC that mines Ethereum
Eugene: Then I started reading about crypto. It was already pretty active, so I'd missed the whole early period. I'm also a gamer, and I was trying to justify building myself a $4,000 gaming machine. Ethereum was about six dollars per token, so I thought: I'll buy a video card that can both play games and mine Ethereum.
Eugene: I built the machine and mined Ethereum. If you've ever mined, you know you can't get into mining without getting very technical. You start configuring blockchain nodes and optimizing gas and so on, and after a couple of months, I got pretty familiar with how it all works.
Paying for the machine
Hummingbot: Was this 2017 or 2016?
Eugene: Late 2016, before the huge run-up, maybe September. By then I understood blockchain technology pretty well. Ether grew to something like nine dollars, maybe a little more, and mining paid for the whole machine plus about 50% profit on top. I was like, okay, I like the game.
ICOs and the 2017 boom
Eugene: Then the ICOs started coming in, and Golem was the first one. It was like startups, but you didn't need to be a venture capitalist to try it out. So I thought, why not? I'd saved some money from my salary as a programmer, so I threw in a thousand bucks.
Eugene: Golem didn't do much for half a year or so. Then I invested in ICN, and that went a lot better. There was a lot more of that, so I started investing in different projects and trading a little.
A $50 token that grew a thousandfold
Eugene: There's a story here. There was a token with a rude joke for a motto. It was for tipping, like when you tip someone on Reddit for a good comment, and it came before the Reddit tokens. I liked it, so I sent about 50 bucks just for fun. Nothing was structured: I went to their Slack, the guy gave me his Ethereum address and I sent it.
Eugene: It grew about a thousand times in half a year or so, and I sold it for about 50,000 bucks. That's crazy, completely ridiculous, but it was 2017. I flipped a bunch of ICOs and traded manually.
The 2018 crash
Eugene: Then 2018 came, and you know how that went for most new traders. I lost a significant portion of the money I'd made. I stopped trading and thought, okay, that thing doesn't work anymore, so what do I do? The psychological load of all of it was pretty harsh, too.
Eugene: By that time, I'd read a couple of hundred investing and trading books, including all the Market Wizards. That would be my first advice: read a lot. Most of the stupid things we do as investors have been discussed already, and there are solutions to them.
8:26 Book recommendation: Market Wizards
Hummingbot: Is there a book you would recommend to people?
Eugene: It depends on where you are in the journey. The first book I usually recommend is Market Wizards, and two or three in the series are quite good. It's a general overview of what's out there: lots of people discussing completely different strategies and ideas. You can't do most of it, but you at least understand what exists, and it opens your eyes a little.
Eugene: There are no crypto discussions in it, but most of the things we do have been mentioned in the book. There's also the usual classic, Reminiscences of a Stock Operator. It's a nice story, a very good read, and it gets you into the mood of how it all works. It's pretty amazing it was written so long ago.
9:29 Book recommendation: Advances in Financial Machine Learning
Hummingbot: I like Market Wizards a lot too. The book I think a lot of quants have been reading recently is Advances in Financial Machine Learning.
Eugene: Yes, I was going to come to that. López de Prado is amazing; he's the main guy. But before you get into that, there's the psychological side. I'd call it very important for understanding how bubbles develop, why people run away from markets and why, for example, liquidity dried up something like 100x in 2018. It tells you what to expect generally.
10:25 Book recommendation: Memoirs of Extraordinary Popular Delusions and the Madness of Crowds
An audiobook about the first markets
Eugene: There's a very nice book that I think isn't available in print; it's an audiobook on Audible. Let me look it up, because it was a very important one for me. Here it is: Extraordinary Popular Delusions and the Madness of Crowds, by Joseph de la Vega and Charles Mackay.
Eugene: It combines two or three books. One discusses how the world's first stock market worked, in Amsterdam. Another covers the most ridiculous bubbles, like tulip mania and the South Sea bubble. When you hear it, you realize how little has changed. They had stock options in Amsterdam centuries ago.
Everything changes, nothing changes
Hummingbot: I worked my first seven years in traditional finance on Wall Street, creating CDOs, MBS and all these things. I agree with you: as much as everything is changing, everything is exactly the same. What we do in crypto today is no different from what I did on Wall Street 20 years ago, or probably what people did in Amsterdam long before that.
Hummingbot: The details changed, but the overall ideas are still very similar, which I didn't expect. Quant traders still need these fundamentals. Learning the art and theory of how finance works is harder than the technical work.
Market regimes
Eugene: Markets go through a lot of regimes. If you developed quant strategies at the beginning of 2017, basically whatever you did, unless you were shorting, would have been successful. But if you don't know how markets and bubbles develop, how people react and what happens after, which I didn't know in 2017, you don't know what to expect.
Eugene: And when you don't know what to expect, you lose like 90% of your money in 2018.
13:32 Starting 3Jane
Getting away from manual trading
Hummingbot: Tell me, how did you end up starting 3Jane?
Eugene: That's exactly how. I lost a lot of money and tried to distance myself from manual trading, so I went into quantitative and mathematical approaches. I'm not a great mathematician, but as a software developer, that route is a little easier. The company was named differently back then. I had three trading companies doing different things, then aggregated them all into 3Jane.
Brute-force optimization
Eugene: I read the traditional material, like Pardo on walk-forward optimization. We did a bunch of ridiculous things, like spinning up a cluster of machines in the cloud, walk-forward optimizing thousands or millions of parameter combinations, picking the good ones and trying to run them. That obviously doesn't work, and you need to read to understand why. We didn't know that yet.
Machine learning
Eugene: After that first stupid bunch of ideas, we went into machine learning. That's where López de Prado comes in; his book is probably the best available on applying machine learning to trading. I read it maybe 16 times before I understood what he actually says, because he's not very clear about his ideas. But they're very good.
Eugene: Basically, he says don't use machine learning to trade. Use it to understand the world and model reality, then trade without it. You need it to optimize other things, while execution doesn't require it and predicting prices usually doesn't either.
Two years of machine learning
Eugene: We spent a couple of years on machine learning. We were moderately successful, but I don't think it paid for the team. Machine learning people are expensive, and I don't have billions in capital, so paying an expensive team from trading profits was hard.
Eugene: Then we got disappointed, and I concluded that simply trying to predict the next bar doesn't work.
Back to classical strategies
Eugene: So we went into classical strategies instead:
- statistical arbitrage
- mean reversion when the relationship is stable
- trend following
- market making, obviously
Read the right science
Eugene: We've been more successful with those, because the people who write the classical articles know what they're talking about, a lot more than we thought we did. So another piece of advice: if you want to be quantitative, read the science. Read Avellaneda, López de Prado, Stoikov, all the people who wrote about the kind of strategies you're trying to use.
Eugene: Preferably the older ones, from 10 to 20 years ago, rather than the new ones. There are nuggets of ideas in new research, but unless you know very well what it's based on, you can't tell whether it's false. We've tested hundreds of papers, and about 300 out of 301 are just false positives that got through with false scientific merit.
18:07 Using machine learning to identify opportunities
How professional firms make money
Hummingbot: People focus on finding a model that beats everyone else's, but professional trading firms make money through connections to many trading venues. Anything that happens on one venue creates arbitrage opportunities, or chances to provide liquidity in one place while hedging in another.
Hummingbot: You're essentially entering advantageous scenarios where you make money whatever the market does. To me, machine learning is best used to identify those opportunities, not to build a black-box golden goose with infinite profits, which never happens. Usually it's human plus machine: the machine flags an opportunity, and a human sets up the bot, runs it for a few days and assesses it.
Ride the easy arbitrage
Hummingbot: That's one of the big lessons we struggle to teach our users. Don't try to create some perfect algorithm. Find some easy arbitrage, ride it for a few weeks until you exhaust it, then move on to the next one.
Eugene: That's one way to go about it. We don't run pure arbitrage, at least for now, because it's very competitive long term. You're right that you can find it if you look everywhere, especially if you don't have too much money and you look in obscure places like new exchanges and decentralized exchanges. They're not perfectly efficient, for sure.
Width and depth
Eugene: You can profit from that for a short time, but it goes away pretty quickly, and you can't trade much money there. So instead we went with a combination of width and depth. We look around and identify opportunities, but we also dig deeper than the plain idea written in an article somewhere.
Eugene: So we have a mathematical advantage in some places and a speed advantage in others. We're a small team and can't outcompete on speed alone. We trade in the low milliseconds, but I know people who trade faster, and they outcompete us.
Where a small firm plays
Hummingbot: There are small firms, big firms and individuals in these markets. Given your size, where do you play, and where don't you?
Eugene: As I said, we look for opportunities. We're connected to most of the exchanges with enough liquidity for us to do at least something, and there aren't many of them, maybe five or so. Then we look at our arsenal of ideas and ask which ones we understand at least a little.
What 3Jane avoids
Eugene: We don't know sub-millisecond trading very well, for example. We can't do it efficiently or fast enough to outcompete, so we don't touch it much. The same goes for extremely complex math and machine-learning predictions. We found some ideas there but haven't been able to launch them.
21:43 3Jane business model
The strategy mix
Eugene: To structure it, we do these kinds of strategies:
- market making
- statistical arbitrage
- directional strategies: mean reversion and trend following
- liquidity mining, which you could count separately, because it isn't the same as market making
Mean reversion needs the right markets
Eugene: The mean-reversion side requires very little math. You need some Kalman filters and a model for the spread, but for experienced people it's very easy. Then you can plot a line where you want to execute, and it more or less works. It's very visual and not complicated.
Eugene: What you need is to see the opportunities where it will work. If you just launch mean-reversion trading on BTC-USDT, it will kill you in a day at our speed. You need markets, or combinations of markets in a stat-arb portfolio, where the relationships you need exist. You find markets that suit your strategies, not the other way around.
Trend following and market making
Eugene: Same with trend following: it works if you find markets that trend. If you trade something like the euro, it doesn't work very well. The same goes for market making. The less volatile the market and the more volume there is, the easier it is. Launch pure market making on BTC-USDT and it will kill you very fast, because the volatility is huge.
Eugene: Unless you do something about that. You can start thinking about catching those waves, which might work, but then it isn't pure market making at the top of the book anymore.
Exchanges that pay you to trade
Eugene: You can also look for relationship-based opportunities, or for exchanges that will pay you to trade. For example, some exchanges:
- subsidize their market makers
- pay rebates
- offer zero fees, so you can execute limit orders cheaply
Liquidity mining as subsidized trading
Eugene: I'd say the liquidity mining programs you run are that kind of thing: subsidized trading. Because of that, it's much easier to provide liquidity in those markets. You can even trade at a small loss, and it's still paid for by the rewards.
Hummingbot: My general advice for individual users is to look for opportunities where you're getting paid to trade, whether that's rebates, liquidity mining or lower fees.
24:58 Advantage of the big players in the market
Small fish, smaller ponds
Hummingbot: Big players like the large hedge funds will always have a huge advantage on the biggest markets, like BTC-USDT on Binance, which will stay very competitive with all the whales there. Those whales are at very high VIP levels, so they pay no fees and even get rebates. Small players have the advantage on smaller exchanges that offer rebates and smaller pairs with liquidity mining rewards.
Hummingbot: It's harder for big players to trade those smaller venues and markets, because they have too much money for the liquidity there. That's why there's a liquidity reward. Small fish go in smaller ponds and earn the rewards that help compensate for the risk of trading.
Team size
Hummingbot: How big is 3Jane now, in terms of people or assets under management?
Eugene: There's no AUM; it's a prop shop. We trade our own capital, which we don't really want to disclose. In terms of people, it's a small team. The size varies a bit, but it's around five people. The smallest we've been is three, and the biggest is 12.
Monthly volume
Hummingbot: Can you share how much volume you're doing monthly?
Eugene: Last time I checked, we did about a billion on Binance last month. It depends a lot on the strategies. If you mostly trade stat arb and mean reversion in a month, it won't be a lot, because some stat-arb trades last days. Pure market making at the top of the book generates a lot more volume.
28:11 3Jane learns about Hummingbot
A month off
Hummingbot: So how did you get involved with Hummingbot?
Eugene: It was after we realized that applying machine learning directly to trading doesn't work very well. I was fed up with it and went on vacation for about a month. My co-founder said, okay, I'll try something different. He does most of the technical work outside research: execution optimization, the platform and other software. He's a really great software developer.
A profitable combination of markets
Eugene: As a software developer, he naturally went to tools that are already available, where you don't need to invent the math, and he found Hummingbot. He tried the grid trading ideas, pure market making and other strategies, experimenting with how they work. After a month or so of that, he came to me and said, look at this.
Eugene: I won't disclose the strategies, but he'd found a combination of markets where Hummingbot's strategies are profitable. I said, okay, that makes sense, and we started applying it. Then I saw inefficiencies, got back into research, and we improved how we model spreads. But he found the general idea just by clicking around Hummingbot and looking at markets.
Math is for optimization
Eugene: After that, we understood that math in trading isn't for the trade idea. It's for optimization. You come up with ideas that don't require mathematics most of the time. You can explain them in words, and they're clear and understandable. You need math when you want to execute them in a way that isn't stupid, because finding something close to optimal usually requires it.
29:40 How does 3Jane use Hummingbot
Hummingbot: Are you using the off-the-shelf Hummingbot strategies, or did you customize your own?
Eugene: We started with the off-the-shelf ones. We traded pure market making for a while and fiddled with the parameters, obviously.
30:08 Customizing Hummingbot strategies
This long chapter moves from custom pricing models to documentation, liquidity mining campaigns and decentralized exchanges; the later article on moving from Hummingbot to custom trading picks up that thread.
Modeling the mid-price
Eugene: Then we wanted to model the mid-price differently. Hummingbot lets you extend some parameters with external files, so we wrote Python scripts with our own math to remodel the mid-price. Hummingbot then traded around it with its own parameters.
Eugene: Next we wanted higher frequency, because off the shelf you get one-second updates. For some things that works fine; for others, you want to be faster. So we contacted your team and asked about the source code and where the parameters are to go faster than one second, and we fiddled with that for a while.
Going custom
Eugene: Then the strategies we were using started to diverge from what Hummingbot provides. We needed hedging on third-party exchanges, taking rebates into account and modeling the typical spreads between prices. It became complicated, and when it gets complicated, you don't have much choice except to go custom. We built a lot of our own infrastructure around it, and we still use Hummingbot for some things.
What would make Hummingbot more useful
Hummingbot: We see a lot of trading firms that have already built their own stacks, or find Hummingbot too basic. We're planning to make it more customizable for these use cases, whether that means more speed or more exchanges, and generally more developer friendly.
Eugene: For us, two things prevent us from using it more. One is speed: we want something like 10 milliseconds, not a second. The other is the way you need to write strategies around it. Right now it's a bit involved. We built some of that, but it could be simpler.
The documentation
Hummingbot: The strategy piece is what we're focusing on right now: simplifying how strategies work and making it easier for people to use the different parts.
Eugene: What I like a lot is your docs. Please keep doing that. The basic strategies you document are what you need to start with, and I learned a lot about market making from them.
The secret of open source
Hummingbot: That's funny, because I think we need to significantly improve our docs. We haven't done a great job of explaining, from a technical perspective, how each strategy works: not just a flow chart, but diving into the code. I think the secret of open source is really documentation, and I'm fully committed to making ours the best open-source docs out there.
Eugene: You read your page on Avellaneda, then you go read Avellaneda himself. Unless you have a mathematical education or a PhD, I'm not sure you'll get any useful information out of the original.
Avellaneda
Hummingbot: There's so little academic literature on market making. We're going to try to talk to Avellaneda and maybe get his advice on how to set up the Avellaneda strategy.
Hudson & Thames
Eugene: Earlier I asked whether you'd seen anything better. I've seen something comparable in only one place: a team called Hudson & Thames, which makes mlfinlab. They took López de Prado's Advances in Financial Machine Learning, implemented its ideas and went further. We used their code for a while, then went custom there as well.
Eugene: Their strong side is the docs. They give you the same information as the articles without the wall of formulas. You need to pay them for the interesting stuff, but it's still very educational.
35:39 Hummingbot Miner
Hummingbot: One more topic: you're active participants on our Miner platform, which is really how we pay for the open-source Hummingbot project to be sustained and supported. What has your experience been so far?
Eugene: It's pretty great. I don't really know what there is to improve except bringing more projects on board that are ready to finance the liquidity you provide. The reporting is good.
Choosing campaigns
Hummingbot: We offer campaigns on Binance, KuCoin and AscendEX. Some pairs have run for a long time, like the Avalanche, NEM and RLC pairs, while newer, shorter-term campaigns on AscendEX have higher yields. How do you pick which ones to join?
Eugene: We don't look much at the yield. We look at whether it can be traded in a way that's profitable for us. Some high-yield pairs are very volatile and can't be hedged, and you need to stay on tight spreads to earn the reward. For some, we figured out how to hedge or stabilize them, or found workable market-making strategies. For others we didn't, and there the yield would need to be huge.
Risk against reward
Hummingbot: So yield alone doesn't capture how much risk you're taking. Either you have the history to know how to handle it, or you hedge it, so you can isolate and capture the reward.
Eugene: You always compare risk and reward, and the reward is basically your yield. When I model it, I consider trading PnL to be zero or slightly negative, which is often true for many pairs. Then what do you lose? For some pairs a lot, for some a little, and some you can hedge profitably. It's a very individual approach: you figure out each pair.
38:54 Decentralized exchanges
Hummingbot: One big topic in crypto recently is the rise of decentralized exchanges, like Uniswap, PancakeSwap and Pangolin. Have you started trading on those venues?
Eugene: I've traded on them for a long time, since before swaps, on EtherDelta and others. There's also Liquid Swap on Binance, which is somewhat of a decentralized exchange. For us, it's complicated. They're not very liquid, and the fees are exorbitant. Most of them only recently added limit orders, which are the only order type we use.
Brains, not capital
Eugene: There are opportunities. I know people who made maybe a million dollars, which is a ton for some people and not for others, by exploiting technical inefficiencies without risking much capital. You stake maybe 10,000 bucks and, if you're smart enough, turn it into a lot more. But that's pure brains. Most of them were blockchain developers who understand the specifics.
Hummingbot: That's consistent with what we see from professional trading firms. Everyone's excited about DeFi, but unless you're deploying less than about a million dollars, it's hard to deploy capital at scale and know you'll make money. The opportunities are idiosyncratic: something happens and you can make 10x, but you wait a long time, so profits aren't predictable.
Trapping the front-runners
Eugene: There's also the usual technology war: people front-run swap orders, and other people try to mess up the front-runners. In one article, a developer made a token with code that, when money went through a certain function the front-running bots used, returned only about 20% of it. He made the token very tasty for front-runners, and they kept front-running it and paying him.
Eugene: That was fun. That kind of stuff works, but you see what skills you need to have.
The home exchange
Hummingbot: In the long run, I think decentralized exchanges will keep getting bigger, and this is a lesson from traditional finance. There were always many exchanges in different countries, but the winner is where an asset is first issued. The first exchange an asset lists on generally becomes its home exchange. In the past, many tokens listed first on Binance or another centralized exchange.
Hummingbot: Now a lot of newer token issuers list first on a decentralized exchange like Uniswap, with the secondary listing on a centralized exchange. If that persists, liquidity for the newer tokens moves to the DEXs, and over time liquidity transfers toward them. It's probably going to take the next three to five years.
What DEXs still need
Eugene: It's going to be a process. DEXs have a lot of advantages. I'm not a big fan of how the current financial system is structured; there's a lot wrong with banks, exchanges and government regulation, and they're not very efficient. DEXs can do a lot better, and in some things they already do.
Eugene: But they're still not technologically sophisticated enough, and the regulators will try to take them away from you, so you need to figure out how to deal with that. On the technical side, they need:
- a better matching engine
- better trading infrastructure, such as order types
- better speed
44:17 Closing
Hummingbot: That's exactly why decentralized finance will last. It's basically the internet, and now that it's on the blockchain, you can't take it away. Whatever any regulator does, there will always be another Uniswap clone that lets people swap.
Hummingbot: Having worked in traditional finance myself, I see what we're building in crypto as a better, freer version of it. I want Hummingbot to be an open-source tool that helps that new world of finance come true. I'm glad we did this, because we share a lot of the same motivations and background: we're here to build a world that's more open and free.
Hummingbot: Thank you very much, Eugene, for using Hummingbot and being part of the community. Hopefully we'll get another chance to do this soon.
Eugene: Sure. Thanks, Mike. It's been fun.
Trader Insights excerpt
On September 21, 2021, Hummingbot published Eugene Tartakovsky 3Jane | Trader Insights, a 10-minute edit of this conversation covering 3Jane's services, optimizing strategies, volatility and finding opportunities.