Extraordinary lessons from a Machine Learning Legend

The financial world already started working with Machine Learning way back in the eighties of the last century. An absolute legend in that field is Jim Simons, who may very well be the best money manager on earth. However, chances are you haven’t heard of him. Many on Wall Street, including competitors in his specialty, quantitative trading, haven’t heard of Simons or of his operation, Renaissance Technologies, either.

Jim Simons was a mathematician and cryptographer who realize that the complex math he used to break codes could help explain patterns in the world of finance. He made billions with his trades. 

Jim Simons started his career in science and in his late thirties entered the world of financial trading. Below we will take a look at his remarkable career in the trading world and his extraordinary results by analyzing enormous amonts of data by using algorithms.

Jim Simons was thinking about how to become rich. While at Berkeley, he bought soybean futures and went to the exchange in San Francisco to watch them being traded. (“They went up! And then they went back down.”) In the late seventies, Simons founded a small investment firm. He tried his hand at currency trading, and then at commodities, but he didn’t enjoy the experience. It was the investing equivalent of wet-lab work. “It was fundamental trading, not systematic,” he said. He felt that there must be a more statistical way to make money in the market. “I looked at the price charts and analyzed them, and they didn’t look random to me,” he says. “They looked kind of random, but not completely random. I felt there had to be some anomalies in this data which could be exploited.”

He hired another mathematician and they began to create models that predicted the direction of currency prices. Simons staffed his hedge fund, the company that became Renaissance Technologies—not with financiers but with physicists, astronomers, and mathematicians. He also invested heavily in computers and in the people who ran them. “If you’re going to analyze data, it really has to be clean,” he said. “Suppose it’s a series of stock prices. 31?, 62?. Wait, stocks don’t double in a day—so there’s an error in the data! There’s all kinds of ways to get bugs out of data, and it’s important, because they can really screw you up.”

There's something called the efficient market theory, which says that there's nothing in the data, let's say, price data, which will indicate anything about the future, because the price is sort of always right, the price is always right in some sense. But according to Jim Simons that's just not true.

So there are anomalies in the data, even in the price history data. For one thing, commodities especially used to trend, not dramatically trend, but a trend, so if you could get the trend right you'd bet on the trend and you would make money more often than you wouldn't whether it was going down or going up. That was an anomaly in the data. But gradually we found more and more and more and more anomalies. It's so overwhelming that you are going to clean up on a particular anomaly because if there were, other people would have seen them. So they have to be subtle things. And you put together a collection of these subtle anomalies and you begin to get something that will predict pretty well.

Well, the system, as it is today, is extraordinarily elaborate. But it's not a whole lot. It's what's called machine learning. So, you find things that are predictive, you might guess, oh such and such should be predictive, might be predictive and you test it out in the computer and maybe it is, maybe it isn't; you test it out on long-term historical data and price data and other things. And then you add to the system this if it works; and if it doesn't, you throw it out.

So, there aren't elaborate equations, at least not for the prediction part, but the prediction part is not the only part. You have to know what your costs are when you trade. You are going to move the market when you trade. Now, the average person will make a – buy 200 shares of something and he's not going to move the market at all, because it's too small, but if you want to buy 200,000 shares, you are going to push the price. How much are you going to push the price? How are you going to – are you going to push it so far that you can't make any money because you've distorted things so much? You have to understand costs, and that's something that's important. And then you have to understand how to minimize the volatility of the whole assembly of positions that you have and be – so you have to do that. That last part takes some fairly sophisticated applied mathematics, not earth-shattering, but fairly sophisticated.

Renaissance has had an unprecedented run with the firm’s signature product, the Medallion Fund, “perhaps the world’s greatest moneymaking machine. For nearly three decades, it has gone up by eighty per cent annually, on average, before fees. Renaissance’s other, bigger funds have done less well. Simons said that this is a consequence of their size: large amounts of money cannot be traded as quickly, and longer-term trading makes algorithms less useful. “It’s like the weather,” he says—the nearer in, the higher the certainty.

Below is a part of a TEDX interview that TED's Chris Anderson(“CA”) did have with Jim Simons (“JS”) about his extraordinary life in numbers. This part is about his trading activities and Machine Learning.

CA: Looking at a typical graph of some commodity, I say, "That's just a random, up-and-down walk -- maybe a slight upward trend over that whole period of time." How on earth could you trade looking at that, and see something that wasn't just random?

JS: In the old days commodities had a tendency to trend. Not necessarily a very light trend, but trending in periods. And if you decided, OK, I'm going to predict today, by the average move in the past 20 days -- maybe that would be a good prediction, and I'd make some money. And in fact, years ago, such a system would work -- not beautifully, but it would work. You'd make money, you'd lose money, you'd make money. But this is a year's worth of days, and you'd make a little money during that period. It's a very vestigial system.

CA: So you would test a bunch of lengths of trends in time and see whether, for example, a 10-day trend or a 15-day trend was predictive of what happened next.

JS: Sure, you would try all those things and see what worked best. Trend-following would have been great in the '60s, and it was sort of OK in the '70s. By the '80s, it wasn't.

CA: Because everyone could see that. So, how did you stay ahead of the pack?

JS: We stayed ahead of the pack by finding other approaches -- shorter-term approaches to some extent. The real thing was to gather a tremendous amount of data -- and we had to get it by hand in the early days. We went down to the Federal Reserve and copied interest rate histories and stuff like that, because it didn't exist on computers. We got a lot of data. And very smart people -- that was the key. I didn't really know how to hire people to do fundamental trading. I had hired a few -- some made money, some didn't make money. I couldn't make a business out of that. But I did know how to hire scientists, because I have some taste in that department. So, that's what we did. And gradually these models got better and better, and better and better.

CA: What role did Machine learning play in all this?

JS: In a certain sense, what we did was Machine learning. You look at a lot of data, and you try to simulate different predictive schemes, until you get better and better at it. It doesn't necessarily feedback on itself the way we did things. But it worked.

CA: So these different predictive schemes can be really quite wild and unexpected. I mean, you looked at everything, right? You looked at the weather, length of dresses, political opinion.

JS: Yes, length of dresses we didn't try.

CA: What sort of things?

JS: Well, everything. Everything is grist for the mill -- except hem lengths. Weather, annual reports, quarterly reports, historic data itself, volumes, you name it. Whatever there is. We take in terabytes of data a day. And store it away and massage it and get it ready for analysis. You're looking for anomalies. You're looking for -- like you said, the efficient market hypothesis is not correct.

CA: But any one anomaly might be just a random thing. So, is the secret here to just look at multiple strange anomalies, and see when they align?

JS: Any one anomaly might be a random thing; however, if you have enough data you can tell that it's not. You can see an anomaly that's persistent for a sufficiently long time -- the probability of it being random is not high. But these things fade after a while; anomalies can get washed out. So you have to keep on top of the business.