Andy Webb: When you're developing your in-house ideas, what tools are you using? MatLab? Octave? SciLab? FreeMat?
Evan Hochstein: None of them, actually. I think because our model of the model creators is very different from what people are used to in the rest of the world. We are a very well-established HFT operation, and we have been for some time, but most of our activity was human-driven and human-created. Although all our trading is electronic, that's not the decision-making method.
The business has decided to embark upon the more algorithmic, the more electronic method of trading, but because that decision came from a manual-trading world, the people who embarked upon it aren't quants, or mathematicians; they are traders. So those models that exist today, they're basically developed by traders who had a trading idea, or a trading experience, based on years of trading. What they've decided is: they have a trading methodology that works, and probably a machine would do it better. So what they've done is create a machine that does what they were going to do anyway.
There was no usage of MatLab or any of those systems. The focus was on the trading systems themselves, on market conditions, backtesting on the systems themselves rather than running theoretical or mathematical analysis. We use third-party common trading platforms, TT or Stellar, and therefore the development of our algorithmic trading was done using those platforms.
Andy Webb: So where those platforms have backtesting facilities built in, you would have been using those to test ideas…
Evan Hochstein: Unfortunately, with the platforms we had, I think we were ahead of them. We had to do our own. But now, this year, Stellar Trading Systems just released their Quantum Server, their algo engine and algo creator, so we'll explore what that does. And TT bought TickIt; they're integrating that into their system, the TT Algo SE. The electronic trading we have taking place today - those platforms didn't exist. We've had to develop to their APIs, write to their APIs, use their delivery mechanisms, their execution mechanisms, but not their logic.
Andy Webb: Such backtesting as you were doing, did you have to write your own backtesting harness?
Evan Hochstein: Yes, we had to. Dot net, C# … yes, and on the infrastructure side, we had to create our own environment. We would create an entire trading environment to enable backtesting and development, a separate development environment to avoid impacting on trading. In one instance, a trader wanted to record an entire market and then play it back applying various scenarios. It sounds very primitive; today there are big, fancy systems, but we were doing it a little earlier. In some cases now, we manage to get a platform that will talk to one of these, and we're using a platform to do it as well. We've evaluated a lot of the known algo platforms, and one of the platforms we've recently looked into, it's from - don't know if you know it, a small company called Deltix.
Andy Webb: Deltix? I've known Stuart Farr since he was at Credit Suisse.
Evan Hochstein: They're amazing guys, good platform, we went out and met them, met them in St Petersburg, went on site, wonderful people. Look - the interesting thing is that we're very different. We're not an obvious Apama candidate, say, we're not even an obvious Deltix candidate. One of the reasons for that is, our traders are geographically diverse. Most of these platforms are meant for use in a LAN environment. That's where they work best. They don't care about the bandwidth they take up, they don't even understand the messaging that's going between the client and the server. But once you hit an environment like ours, where people are WAN-based and bandwidth is a bigger problem than latency, these platforms don't work so well.
Andy Webb: Fascinating…
Evan Hochstein: We were happy with Deltix, they were very accommodating, ready to learn about this market and make any changes. In the Deltix platform we've developed some tools. Right now, we've used the Deltix platform to develop mainly what I would call indicative tools. By that I mean a human being is still the trader. He still decides whether to buy or sell - rather than being the pilot on a 787 and just doing autopilot: on/off. He trades, but the indicative tools tell him - hey, that scenario is coming now. What you've been looking for based on your settings - here's what's going to happen, when it's going to happen, get ready, shoot. Indicative tools are extremely important, maybe even sometimes more important than the algorithmic tools, especially in businesses that are not latency-sensitive.
Those people who look for split-second trading, high frequency, colos at the sites, I don't think they're trading futures or equities. They're trading time. That's all they're running after. Obviously, in my world, it's my job to ensure that once the trader wants to make a trade, it happens as efficiently as possible. If I can host the execution aspect as close as possible to the exchange I will, to ensure that he gets executed right away. But he's not trading time. He's trading a market idea. Therefore, he's not sensitive to the timeliness of it. That's why I think indicative tools are very, very important. They allow the trader to test the market, get a feel, test his own scenarios, and he's still the trader.
So we've developed a bunch of indicative tools with Deltix, and we've given them out to around 300 of our traders, and we have a team that's developing trading tools on the same Deltix platform. Again, Deltix is a platform for them mainly to develop in and to distribute with, but it's going to shoot the trades out through an ISV, Stellar or TT, connecting through the APIs on the back end, et cetera. We don't develop our own execution engine and we're not connected directly to the exchange. We're still using ISVs for the physical trading.
Andy Webb: When I last spoke to Stuart a few months back, one of the issues he was looking at was the fact that they don't have native data themselves. Given the number of connections you've got to various exchanges, you've got access to live market data from the market gateways. How do you handle your data management so that the guys building, testing models have got easy access to everything they need?
Evan Hochstein: It's interesting. We thought that we were so fortunate - we have twenty-eight exchanges coming into our systems, we have access to everything, we can just farm data. That turned out to be totally unsuccessful. I think people spend more time cleaning data than finding it. Today there might be tools that do a better job. What we've found is, we subscribe to a Reuters historical data feed, and that's what we use. We didn't have to buy any of the known tick-data databases or tick-data farming systems, because Deltix handles it actually quite well. Deltix has its own, let's call it tick-data database feature: we feed into it a Reuters live historical data feed, and Deltix does the rest. Actually, one of the reasons for using Deltix as a platform for developing tools was these features: first, its ability just to take tick data. I didn't have to buy a service. Secondly, its ability to have backtesting built into its platform.
Andy Webb: I reviewed Deltix a few issues back, and one of the things that impressed me with the tick data and the way it managed it was the speed and the efficiency of the bar-building algorithm in there. If you didn't want ticks - you wanted constant-volume bars or whatever - it built the bars extremely quickly from the tick data, much quicker than I've seen elsewhere.
Evan Hochstein: Our people are very much bar traders and candle traders, and that was very important for us. We explored all the main guys, and maybe one day I'll need them, but right now, for what I need, this works perfectly for us.
Andy Webb: You mentioned TT as one of your market-access providers. Who else are you using, or you're likely to use, on this automated-trading project?
Evan Hochstein: TT is one of our providers of market transaction ability. We use their front end to trade, and we have their back-end servers. We have our own exchange connectivity that we bring it, et cetera, but the trading goes through them. We have TT, depending on the markets; and unfortunately, TT doesn't cover all the markets in the world. We have Stellar Technologies, superb company, amazing company, yes it's young, it's small, but I think their product is way, way advanced. Stellar have the ability to allow anyone to plug into them with multi-user algos. We use also Pats. On markets that neither TT nor Stellar have touched, mainly Asian markets, Pats is already there. What drives our usage of these platforms is the markets that they are in, and of course their features and the user requirements. We'd like to accommodate any trader to do what he's familiar with and what does it for him. Our three big ones - Pats, TT and Stellar.
We did a fashion show, whatever you call it, for a bunch of the algo platforms to our traders, to see what they would find interesting, and I'd says that the Stellar Quantum Server looked a lot more robust, a lot more powerful, a lot more feature-rich, easier to work with.
Andy Webb: We looked at Stellar about nine months ago for a review, and you can get into it at a much lower level, you can do so much more with the thing than you can with the others, get much further under the hood. The one thing we did find - this was nine months ago - the documentation? The basic nuts and bolts, I'd agree with you, pretty good. But… ?
Evan Hochstein: Yes, but to their credit they're willing to talk to you and help you out on that. I'd also say that the Stellar front-end programming would require you to be a bit more of a technical guy, more of an engineering background than an economics background to understand it and work with it, while at TT it's the other way round - really drag, drop, click, put here.
One of our requirements at the fashion show was - to anyone we spoke to - will you accommodate plugging in an external data source into your algo engine to enable algos making decisions based on another data source as well? The purpose of that question was - we had in mind the intelligent news feeds, Dow Jones, now the Deutsche Bourse. Stellar had no problem with that at all.
The big fear with totally automated and algorithmic trading is risk management, not only from a P&L perspective, but also legal and regulatory. What is the thing doing when the guy goes to the bathroom? How quickly is it going to shoot out thousands of messages? What do you do to mitigate that? What do you do to ensure that this machine gun doesn't go crazy? Look, we're seeing this as an extension to our existing human click-trading. We're already electronic, connectivity is already there; how do we make the traders better? One is through indicative tools, and the step on top of that is fully algorithmic, where the trader is more of a manager of an autopilot engine.
Our main thing is not to trade time. We're looking for ideas in the market. We have a bunch of different guys with different ideas. They came from the market, they're market users, they have ideas and we're programming their ideas.
Andy Webb: Evan, thank you.