Remember me

Register  |   Lost password?

 

 

SUPPORTED BY
Nvidia
GPUs, Monte Carlo Simulation and Kooderive with Professor Mark Joshi - February 25-27th, London, UK - Further Information


The Brain Man: Richard Peterson, Sentiment Analysis and MarketPsychData

Thu, 10 Jan 2013 14:30:00 GMT

Back in September of last year I was fortunate enough to attend the Battle of the Quants event in London, a veritable schmoozefest for some of the brightest folks in the Quant world, stuffed full of rising stars with a few legendary old timers thrown in for good measure.

On Day 2 I was pleased to grab 20 minutes with one of those rising stars, Dr Richard Peterson the CEO of Market Data Vendor, MarketPsychData, which puts sentiment analysis front and center of their business. We cover quite a lot of ground in the interview, from Richard's unusual background as a Quant and a Psychiatrist, to his early experiences working in neuroeconomics and text-mining, through to the history and development of his firm, and what to expect in the future of this fast-moving field. Enjoy.

 

Jacob Bettany: I'm very pleased to be here at Battle of the Quants with Richard Peterson who's the CEO of MarketPsychData. We're going to talk a little bit today about sentiment analysis, along with the technology they've developed, where they were, where they are, and how they’re planning to take the business forward.

Thanks for joining me Richard. First of all I thought it might be a good idea for you to introduce yourself as I think you have perhaps one of the most interesting histories as a Quant I’ve come across.

Richard Peterson: Sure Jacob, no problem. Thanks for opportunity to talk to the MoneyScience users.

I think what most people find interesting about the story I tell is that although I'm in sentiment data now, before that I was in psychiatry - I was a psychiatrist - and before that I was a Quant, a Futures Trader.

I actually got started in this business because I did an electrical engineering degree and part of my senior project was developing quantitative learning models that learned patterns of price in volume data and obviously used that to predict future price activity.

We saw 2 things in that data that were worrisome. 2 aspects of our predictive models concerned us. One was that we saw alpha decay. In fact the alpha that we were finding - the performance - was diminishing in each model over time, so as we approached the mid nineteen nineties’ when I was doing this research, our models had shorter and shorter lives.

The other finding that disturbed us was that our models only worked successfully in periods of extreme emotionality, high fear or high optimism. So firstly I became very interested in understanding why people individually make these mistakes which drive market prices. Secondly, as a Quant, I was asking how do I conceptualise or quantify this behaviour so that I can extract it as a continuing and robust source of alpha, in financial markets and other places.

I went on to medical school from Electrical Engineering and trading as a Quant, and I studied for my senior project the social psychology of bubbles and panics. From there I went on and did a psychiatry residency and did a post-doctoral program in neuroeconomics at Stanford, where we looked at how people make investment decisions on a neurological level. What are the brain activations, for example, that predict investment mistakes or non-optimal financial decisions?

JB: It certainly doesn't sound like a particularly normal route for a Quant. Was it something you did with a plan in mind, or was it something you fell into?

RP: Since I was a kid I was always very interested in investing and it turns out my father was a professor of finance. He gave me some money to invest as a child and I would buy the Fortune Magazine Top 10 Companies and universally every year I would lose money on those purchases, and at the same time I would watch him buying bankrupt airline debt and tripling his money on that. I would say, “Dad, what wrong? I'm buying what I'm supposed to buy and I'm not making any money” and he would say,”well, you've got to buy what nobody else wants”.

And that's when it dawned on me that the trick to successful investing is understanding what people want, and what they do not want. As Ben Graham says, "We buy from pessimists, we sell to optimists" - so how do you quantify that statement in the markets? That became a very enduring interest of mine.

JB: So how did you get from neuroeconomics back into finance? It was obviously a completely natural step at that point!

RP: Oh, Completely Natural! Doing the neuroimaging research we noticed that emotional primes we gave to subjects would change their decision-making from optimal to non-optimal, or non-optimal to optimal - meaning mathematically optimal in the Bayesian model. So we found for example if we offered someone $10 guaranteed profit, or we said on the flip of a coin you could win $30 if it lands on heads or zero if it lands on tails. So the expected value of the coin flip of course is $15. We found on average the groups that we tested split evenly 50-50. 50% would take the guaranteed gain, 50% would take the higher expected value and risk.

What was interesting was that some of my colleagues used emotional priming. Before the question was asked they'd show people a face and say, label this as a man or a woman - and the face would have an emotional expression - maybe fearful or angry, maybe happy, or maybe neutral. What they found between the conditions was that with the fearful and the happy face there was a 30% variance in risk taking.

More importantly after the study when the people were asked, “Did that face effect your decision making?” universally they said, “No, why would that affect me, that was a just a random face”. And when they were asked on a scale of 1 to 7, 'how do you feel?' – before and after the experiment they would rank themselves a 4 on average for both. The priming had affected their decision making dramatically, but it hadn't affected how they felt, it hadn't consciously changed what they thought about what they doing. So I became very interested in how text information and news were priming people emotionally, to make good or bad investment decisions.

In 2004 we started a text mining project. Essentially we created a search engine technology and started downloading all the content we could get from the internet; social media largely, some new media, regulatory filings, SEC filings, earnings and conference call transcripts. All of this information we downloaded and quantified into scores; The amount of fear expressed, the amount of danger expressed, the amount of optimism, joy etc.

Right now we have over a thousand variables though we initially started with about 40 and we began to do quantitative studies and back testing on this data and we found that in fact they do predict stock prices systematically. Not in an obvious way - linear models don't work on this data, but other models did work and we found there was really good alpha in there if you knew where to find it.

JB: The amount of data is obviously proliferating enormously, so I guess in order to bring in as much data as you can and analyse it the technology requirement must be significant?

RP: Right, you pretty much universally have to be on Hadoop and in the cloud generally so that you can scale up for large jobs. We can be renting up to 3,000 servers at any one time to process text. You first have to analyse whether the text is relevant? What are the entities? Are the references to the entities relevant? What are the concepts to be extracted? What is the grammar? It's a 12 step process in our software which is very computationally expensive - and takes time. Even with a lot of speed-boosting technology, it can take us 250 milliseconds to analyse an article because we're looking at so many factors.

JB: Are those thousand factors you mentioned applied to everything? I mean how does that break down?

RP: Okay. So one of our factors is 'Anger at management'. So if someone says, "I hate the CEO" we'll pick that up and score it, but we won't score 'Joy' because it's not in that statement, there is no optimism in that statement. On the other hand if someone says earnings are going to exceed expectations then we'll score 2 meanings, we'll score 'optimism' and we'll score 'earnings expectations positive'. So we score based on what the content is, and we've found over a thousand relevant concepts in business text.

JB: So that initial text-mining project developed in MarketPsychData?

RP: Right. Through trading actually. In 2006 we began a paper-trading portfolio and it was very expensive to design the software - it takes a lot of money - so we started paper trading it to show there was value in it and earned 20% on paper in 18 months. We raised a small amount of capital from investors and launched, in September 2008, a market neutral long-short hedge fund and we traded that for 2 1/4 years - and made about 27% overall - but importantly in the first year, during the crisis, it was up 40% putting it in the top 1% of funds.

As the market shifted, the predominant sentiment changed from absolute hysteria to actual green shoots of optimism, we found that our models didn't work as well. Then there was a liquidity squeeze, we couldn't raise a lot of capital, we lost our CTO, we lost our software developers, so we had trouble pivoting at that point. We continued to run with our old strategy, lost a little bit of money and then we closed it down. We still outperformed substantially but found that even with the liquidity squeeze we could raise money if we spun off the data business as a technology business.

We did that so that we could bring on the talent we needed to build it out again, to continue to improve it, and that's how MarketPsychData was born, as a spin off from the hedge fund structure.

JB: I understand you have a big deal with Thompson Reuters. Is that the first major appearance of your technology in the marketplace?

RP: That's right, the raw data is being distributed through Thompson Reuters but this data has an incredible amount of value in different ways once it's interpreted for users. So Quants interpret it for themselves, we sell the pipe through Thompson Reuters and the Quants take care of the rest. They test it, find the alpha they want, make sure it's uncorrelated with their other factors and then they'll use it in the models.

But we have other businesses. We also do investment relations consulting, In one example we worked for a Fortune 100 company and found that whenever the CEO said certain types of comments the stock price dropped, when they made other comments the stock price would rise. Of course there are obvious factors, talking about earnings, if your earnings are down your stock's going to drop, but when we controlled for the obvious factors we found certain psychological subtleties in the conversation, for instance the number of times he talked about new innovations in his company, that drove the price up; the number of times he talked about 'exciting future possibilities; that drove the price up; whenever he talked about uncertainties or worries - which is unnecessary to talk about - it drove the stock down. We showed this to them and they have refined their communication strategy since.

There's also quantitative products like stock screening tools for long only value funds, something we're working on now - it's not out yet - a sentiment screening tool. It turns out that just like Ben Graham said, you buy from pessimists. But how do you know where the pessimists are? Well now we can quantify that, based on forward looking negative statements about a company, so we have a stock screener where if you buy the bottom decile and short the top decile of the 6000 US stocks, you earn on average about 15% annual, and it's a market neutral strategy, a decile strategy. When you go to the extremes are very nice, so it's pretty interesting.

Oh, I didn't mention, we also have some predictive research products. We have trading signals right now that are idle because we've closed the hedge fund and those have been performing 30% annually with very conservative transaction costs - a 40 basis point round trip assumption - on mid- and large-cap US equities, so a very good performance and we're able to sell those now as a research product.

What I found in running a small fund; the long term potential was terrific but in the short term it's a very tough time to have a small fund and to raise capital. Hopefully the environment will return for small funds but right now we find there's a lot more interest and brand building to be done in the data business - even though we're selling our alpha.

JB: This does seem to be a business sector that's growing very quickly, how do you distinguish yourselves?

RP. There are a lot of start-ups in sentiment and analytics, for sure. What differentiates us is that we have a trading track record. We're the only company that can show audited performance and we also have emotive data, so no one else has emotion as we capture it - and we're from finance, we're traders fundamentally - and most of these companies don't have traders involved. We actually understand the markets and which data is useful.

JB: With the proliferation of companies bundling into this space, how do you see it panning out over the next couple of years?

RP: It's certainly a tough space to be in. A lot of companies are pursuing the raw data for Quants angle and there are very rigorous requirements for data quality from quants, as well as big volume requirements. They want a lot of data. They want to know every source it's coming from, they want to be able to tease it out, which is what I did myself, I get that. But many of these companies have perhaps a feed attached to twitter only - and they're selling it for the price that we're selling every source, 2 million websites being screened. So it's a difficult space. If you're a good salesperson, you can perhaps sell a little bit, but to survive as a company is very challenging.

JB: Will there be some consolidation among these companies, or will they just fail?

RP: I believe the first company in the space was in in '98 or '99 and almost every company that's tried has not succeeded. Some have succeeded in the brand space. Like Umbrio was bought by JD Power, Buzzmetrics was bought by Nielsen, Radian6 was bought by Salesforce.

In the finance space, amazingly to me, it's not lifted off as I would have expected, given the trading results that we see. The reason is - in my opinion - that Quants have learned the wrong statistical tools in graduate school. They've learned linear models, and linear models don't work on sentiment data - or they don't work well. I see people twist and turn them in all contortions to get some alpha, and you get a little bit, but the big alpha is in other types of modelling which people have not learned in their Masters of Financial Engineering.

JB: Well that was very interesting Richard. Thank you very much your time. I'll look forward to speaking to you again.

, , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,