A hot new hedge fund is based on smart computers picking off dumb ones

caption
Manoj Narang.
source
Brendan Smialowski / Getty Images

Manoj Narang has been using data and technology to invest for nearly two decades. Now, he’s launching a $1 billion hedge fund that combines computer-driven investing decisions and high-frequency trading with data on stock market patterns. In the industry, these are called quant and alternative data strategies.

Hedge funds have been adding so-called alternative data – which gives them a read on everything from spending patterns to the impact of weather on retail sales – to their investment analyses, and hiring people who can sort through it. The market for this data is expected to double in the next five years.

Narang’s MANA Partners, which incorporates this kind of data with high-frequency trading, could be one of the biggest hedge fund launches next year.

It opens January 3 with about $1 billion under management at a time when most startups are struggling to raise money.

Business Insider caught up with Narang to chat about the latest trends in high-frequency trading, alternative data, and “quantamental” investing – which combines fundamental stock analysis with the use of computing power and big-data sets to test hypotheses. Narang previously founded Tradeworx, a high-frequency trading firm.

What follows is a lightly edited transcript.

Rachael Levy: Tell me about MANA Partners’ strategy, which seems to be a combination of high-frequency trading and statistical arbitrage.

Manoj Narang: First of all, the fund is not just about high-frequency and stat arb. Our firm is kind of set up to take advantage of these long-standing secular trends. Essentially, as the quant trading space gets more and more competitive, I see more and more convergence happening between previously disparate or segmented areas of the space. So the ongoing confluence between high-frequency trading and stat arb is one of those. The ongoing confluence between fundamental and quantitative investing is another. And then across asset class trading – like, for example, across options and equities markets – is yet another. So we’re trying to position ourselves to take advantage of all these sorts of convergences.

Levy: What’s the advantage to that kind of strategy?

Narang: So the majority of high-frequency trading firms out there are proprietary trading firms. They trade their company’s capital or their founders’ capital or some combination of those two things. They have a pretty high aversion to risk. That presents an inherent limitation to them. Being better capitalized than other high-frequency trading firms is definitely an advantage to us that we plan to leverage.

And more importantly, it’s a pretty compelling combination for investors because high-frequency trading strategies tend to do very well in volatile markets, whereas more conventional quant strategies tend to suffer in volatile markets because of the rampant liquidation pressures that go along with elevated volatility.

The overcrowding of stat arb is a big risk factor. The demand for quantitative strategies has gone way, way up, but on the other hand, the flip side of that, is that the conventional strategies that people run have limited capacity, and they’re now overcrowded. And because they are overcrowded, they are subject to these periodic, protracted bouts of liquidation that tend to coincide with high volatility, and that happens to be exactly the right environment for high-frequency trading to prosper. The two strategies tend to diversify each other during periods of very high volatility.

High-frequency strategies tend to do well in all markets, but exceptionally well in volatile markets. They just have very limited capacity.

Levy: Why have the stat arb angle?

Narang: Stat arb – I think it’s an extremely exciting space. I think we’re at the cusp of a new renaissance in statistical arbitrage. There’s been a huge explosion of the amount of data that’s available in quantitative strategies, as well as recent advancements in nonlinear modeling techniques, like statistical learning methods. I think the combination of those things, as well as the rise of cloud computing, the confluence of those trends has created – I think we’re at the dawn of a new era, if you will, not to be trite or anything, in statistical arbitrage.

source
Mario Tama/Getty

People who pioneer new approaches in stat arb are likely to do well in the coming years. A lot of that has to do with developing strategies that are not overcrowded because they use different types of data or strategies.

There are countless stat arb strategies that are using the same inputs, analysts’ inputs, company filings, or corporate announcements, what have you. Those things have been arbed to death over the last 15 years, and there used to be a lot of alpha. Now there’s still alpha, the market still has to price in new information, but competition has really eroded the competitive advantage. So you need new approaches to be competitive now.

All of this newfound computational power and all this newly available data and newly available methods have made it possible to create new categories and strategies that did not exist before. In five years I think you’ll see, potentially, a new set of players in the market that have been able to get traction.

Obviously, the D.E. Shaws and the Renaissances and those kinds of firms – you know, the Two Sigmas – that have been there a long time are likely to be there as well, but I think that it’s possible for some new firms to get some pretty strong footing by pioneering some new methodologies.

caption
Two Sigma’s logo.
source
Screenshot via Vimeo

Levy: Where will you be getting data? From outside vendors?

Narang: It will be a combination of things. There are certain areas, because of the nature of our business, where we have a proprietary edge in data collection. So for example, we have a very strong infrastructure for both high-frequency trading and for options trading, and that allows us to collect and manipulate data that arises from those markets very effectively.

And those are two of the biggest big data problems that are out there. The options market generates more than 10 times the data as the equities market does, and same with micromarket structure. You’re talking about over a dozen exchanges with their own direct feeds, and you have to properly collect the data and synchronize and time-stamp it. And it’s just a massive amount of data. That kind of data is, we feel, very differentiated. There aren’t that many firms that have the ability to incorporate signals from those markets into their strategies.

Levy: This is data that comes naturally from the markets trading. So no one has tried to collect it?

Narang: People collect it, but it isn’t really commercially available. Unless you’re a high-frequency trading firm, there’s no reason to collect this data.

Even with the options market, certainly you can collect the OPRA feed, but it’s just a massive amount of data to process.

Levy: Can you give an example of what you can do with this data? What would you end up using that data for?

Narang: The options market definitely contains information about the equity market. The options market is all about how investors are pricing in risk. That’s why there’s this notion of implied volatility. Implied volatility is essentially how investors are pricing their estimates of risk into the market. There’s definitely information there about the underlying stocks. We think that that is very fertile ground for alpha.

caption
IBM quantum computer scientist Jerry Chow
source
IBM

This just goes to the broader picture that I painted for you about convergence. It’s very difficult to have that capability if you’re not already an options trading firm. You just wouldn’t have that infrastructure you need in place to collect and process that data, and the same goes for market microstructure. Unless you’re a high-frequency trading firm, you simply don’t have that data because you’re not subscribing to the direct feeds of all the exchanges and transporting the information over high-speed networks.

You can certainly buy historical feed data, but you can’t replicate the effective transporting that data around to get a unified picture of the market. You can only do that by self-time-stamping and self-collecting the data.

The data we’re talking about is hard to come by, and unless you’re set up to trade multiple asset classes with multiple styles, you don’t really have the ability to do these kinds of trades.

Levy: That’s really interesting. It almost seems like this is an access trade in a way because not everyone has access to this information by the nature of how they’re set up.

Narang: Everyone can have access to it. There’s no information here that is privileged. The data is just very expensive, and unless you’re in that business, it doesn’t make sense to purchase it.

Levy: What do you think of the idea that the strong signals in data have become too noticed, essentially? That some of these quant firms are looking for weaker rather than strong signals?

Narang: I would characterize it differently. There is so much data out there and there are so may people now with a data science pedigree that if you take a purely data-driven approach, there are two pitfalls there. The first pitfall is that other people are likely to find the same thing as you are, because everyone knows the same analytical techniques for the most part. And the other pitfall is that there’s just so much data and the search space for models has such high dimension that you’re likely to find strong signals in the data that are just really spurious in real life, just from overfitting to the data.

source
WOCinTech Chat/flickr

What I really favor are more structural approaches that reflect an understanding of how markets actually work – both at a macroeconomic level all the way down to a microstructure level, including a fundamental level in between those.

So I think an understanding of markets and securities from a fundamental level, from a macroeconomic level, and from a microstructure and policy level is crucial to building models. To me, the ultimate way to build models is to basically automate the process of discretionary investors.

That’s where this whole quantamental opportunity lies. More and more, you’re going to see machines replicating the decision processes that human beings engage in when they trade securities. As things stand now, humans still are responsible for the vast majority of capital allocation, and that’s because long-term investors control the largest amount of capital and it’s controlled by human beings. I’m not talking about passive stuff, which is mostly automated. But when it comes to active management, long-term capital allocation decisions are made by human beings.

Even when it comes to indexes and how you allocate to indexes … it’s pretty much done by human beings as well. So the way to really build a compelling quantitative strategy is to understand how humans make those decisions and attempt to replicate those behaviors, and anticipate those asset flows and those allocation decisions and investment decisions. Ultimately, that’s what moves markets.

People have this misapprehension that quants make money because they’re smarter. But that’s not how you make money in the market. You have to make money by doing the same thing as the rest of the market, only doing it first before the rest of people commit their capital. That’s the only way to make money when you’re automated or discretionary.

source
Flickr/Rachel Johnson

The same thing goes for machines. To me, the most fertile ground for building quant trading strategies that are profitable is by anticipating, is by orienting those strategies to have a very strong structural component that understands how human beings make decisions. So that means incorporating fundamentals into your model. I don’t mean just fundamental data, but fundamental reasoning, macroeconomic reasoning, and reasoning about market microstructure. The more human knowledge you can bring into the domain, the more likely it is that that data will assist you rather than confound you.

I would characterize that as a top-down approach to markets rather than a purely data-driven approach.

Levy: So some of the stereotypical alternative data is like credit card exhaust data or Foursquare foot traffic data at certain stores. Is that “traditional” vendor data something that you will be incorporating then, or is it more in-house exhaust data, so to speak?

Narang: There’s no data that we would not be interested in. But we are by and large interested in data principally if we can figure out how a human being would model that data, rather than just a purely data-driven approach. A typical quant will trade a data set, and not really try to understand what the data means, and just fit models to it. And that’s not really the approach we want to take.

source
Business Insider / Pamela Engel

The whole notion of quantamental and this convergence between man and machine – it’s more than just looking at next-generation data. It’s not just about looking for more and more kinds of data. That’s super important, and it’s a key component, but most people doing that don’t have a strong grounding in how markets work, or how companies work, or how market microstructure works, or about how the macroeconomy works. And so they’re relegated to taking a truly data-driven approach to mine those data sets for alpha. And that’s going to lead to lack of differentiation and overcrowding. There are too many people who can do that. There’s no barrier to entry to do that.

The real barrier to entry is still knowledge of how markets work, how companies work, knowledge of how fundamentals work, knowledge of how macroeconomics works, knowledge of how microeconomics works. And if you can bring that knowledge to bear, you can harness that data in a more structured way, and build models that are less overcrowded, and have considerable differentiation. Also, it’s more likely to work from the following fundamental perspective – which is models should only be expected to work if they properly anticipate what other people’s models are going to do. To do that, you have to understand how people who are committing capital think.

Levy: So it sounds like humans are always going to be at the helm.

Narang: Humans aren’t making decisions other than how to build the models. It’s a purely algorithmic trading firm. It’s just when humans are making the models, the humans that are making the models happen to have a strong grounding in traditional investment styles. So what we’re trying to do is automate traditional investment styles much more so than just taking a purely data-driven approach. These are two fundamentally opposing schools of thoughts of how to use data. It’s bottom-up versus top-down.

source
Thomson Reuters

Levy: In terms of the future of quantamental then, do you think there is a general push in using quantamental strategies in a top-down or more of a bottom-up approach?

Narang: The majority of people who do quantamental will take a bottom-up, so a truly data-driven approach, because data is readily available and the skill set to analyze that data is also widely available. There are plenty of people with a data science pedigree, and many of them have access to data. So it stands to reason that that’s going to be a pretty widespread approach. But that approach is going to be highly commoditized. It will be far less common for people to take a top-down approach because it’s much harder to do. It’s much harder to find people who come from the traditional investing universe who also have the ability to put together quantitative strategies. Quantamental is about more than just data – it’s also about methodology.

Really, if you want to get at the heart of what quantamental is, it’s not just looking at fundamental data. Quants have been looking at fundamental data forever. That’s nothing new. What really it’s about is about taking the decision processes of discretionary fundamental investing firms and automating them to the maximum possible extent by systematizing the decision process.