Bill Reid

Artificial intelligence, neural networks, chaos theory — these are all part of the discussion when talking forex with Bill Reid, manager of the Algorithmic Currency Fund (www.algorithmictradingadvisors.com). But along with these concepts, Reid shares some pretty basic ideas about trading currencies.

“I often tell people forex is the equity of one country compared to the equity of another country,” he says. “How that country is positioned in the world market determines its relative equity, and the value of this country’s equity at this instant makes the market very interesting. That makes it a fun market to learn about and participate in.”

Reid’s Fairview, Texas-based fund has posted 12 winning months out of 18 since its launch in September 2005, and its performance — while off its explosive early pace — has nonetheless outperformed both the S&P 500 index and the general universe of professional currency traders — which, as represented by the Barclay Group’s Currency Traders Index, posted negative annual returns in both 2005 and 2006. The Value Added Management Index (VAMI) graph compares the growth of a $1,000 investment in his fund to a $1,000 investment in the S&P 500. Reid currently manages $10.7 million through the fund.

Reid, 65, got involved in trading after taking early retirement in 2004 from IBM, where he worked for a decade in the artificial intelligence (AI) division, developing AI for a wide range of companies, including Boeing.

“All the jet engines use artificial intelligence to determine how well they’re doing,” he says. “After I took early retirement, someone asked me if I ever thought about trading commodities. It got me thinking that artificial intelligence was [mature] enough to be applied to trading. I started talking to people about the application of AI in the financial markets.”

For Reid, who has a master’s in computer science, the transition into trading turned out to be very much an extension of his education and varied professional background. He holds seven patents, including a device for landing on other planets that he developed for NASA early in his career. Forex represented a new challenge.

“I’ve always been interested in new things,” he says. “I wanted to see if we could develop a unique approach that could be patented in this crowded market space. I think we’ve accomplished that.”

Reid trades 10 currency pairs (EUR/USD, GBP/USD, NZD/USD, AUD/USD, GBP/EUR, USD/CHF, EUR/CHF, USD/JPY, EUR/JPY, and USD/CAD) using a trading model built around the concepts and processes he developed in his previous career. When he began researching trading, the Massachusetts Institute of Technology (MIT) was a prime source of ideas.

“I worked with MIT on several financial topics,” Reid says. “They’ve done a great deal of original research in the subject of science application in financial markets. A key part of their work is a wonderful book, Nonlinear Dynamics, Chaos, and Instability explaining how to analyze financial markets with chaos theory. Their analysis indicated the only market that is chaotic is forex.”

We: What does that mean, exactly?

BR: MIT’s mathematical studies found that treasury bonds are random, the stock market is correlated, and forex is chaotic, which refers to something that looks random, disordered, or irregular and has the potential of having some underlying order, if you can figure it out. Weather forecasting, epidemics, and the creation of new planets and galaxies, for example, are chaotic. The forex market is unique because it is bounded by central bank actions on currencies. If a particular country’s currency rises, other countries cannot buy their products. If a country’s currency falls, they cannot buy othe rcountry’s products. This chaotic behavior then applies to individual currency pairs. The bounded action of currencies probably contributes to why, mathematically, the forex market can be shown to be chaotic and exhibit the ability to be traded successfully both short and long term.


We: What attracted you to a market you’d classify as chaotic over one with a presumably higher degree of order, or correlation, such as the stock market?

BR: Chaotic means you can find the order both in the short term and the long term, so you can trade it both ways. The banks trade it long term, but we trade it short term. This short-term, long-term differentiation is not applicable to a correlated market, such as the stock market, because price moves in stocks can result from either general market movement or individual company success. Stock prices cannot differentiate between the market or specific company value driving the price. This is why equity hedge funds require a long funds lock-in period and why all financial advisors advise clients to look at mutual funds from a long-term perspective.


We: In regard to the long-term, short-term aspect of chaotic behavior, are you referring to fractal (see “Chaos theory”) properties — similar patterns or behavior appearing on different time frames?

BR: Fractal patterns certainly exist in 15-minute samples, one hour samples, and four-hour samples. In the longer term, a country’s economic position on balance of trade will be a major factor. Recently, at the Texas Hedge Fund forum, we showed that if you held the 10 currency pairs we trade for nine months, invested 10 percent of your funds in each pair and traded at 50-percent leverage, you would have made 17.7 percent annually — if you had picked the right starting position. Balance of trade positions move slowly.


We: So, how does AI and this concept of market behavior manifest itself in your trading program?

: We start with a large number of inputs around 50 — that go into a neural engine. Twenty-five percent of them are conventional indicators, such as accumulation-distribution, MACD (moving average convergence-divergence), the Relative Strength Index, or stochastics. Another 25 percent are non-linear indicators that have logic to detect the initial conditions. [These are essentially] indicators that look at what the initial conditions were when a certain pattern formed in various places throughout the data history we’re analyzing. It produces a forecast based on that information, and high predictability makes it tradable. Another 50 percent of the inputs are chaos-developed price patterns. These patterns can be individual or interrelated.


We: As in the relationship between two different currency pairs?

BR: Right. The model treats signals as geometric objects — for example, the double peaks that often occur in currency pairs. But the characteristics vary with the frequency, or time characteristics, of the prices. If the rise to a peak or the fall to a valley is short — that is, higher frequency — the likelihood of a double peak higher than the first peak, if it occurs at all, is greatly reduced. We developed a tool to look at the price history to develop these “chaos” patterns, and that actually came out of a model that MIT developed for analyzing chaotic markets.


We: You mentioned you designed a neural network to process all this information. Can you describe what that is?

BR: A neural network is an interconnection network with the ability to train a group of inputs to match a desired output signal.


We: Are the inputs the

different pieces of market information you look at? Is the output hopefully a profitable trade signal?

BR: Yes. We developed profit “fitness” criteria that a genetic algorithm (see “AI and neural networks,” above) uses to produce a profitable output signal. For example, one criterion is how far ahead on the signal we can look to see if this was a good trade. Generally this look-ahead is 10 to 15 periods. Looking ahead 10 to 15 periods generates an output signal that is 100-percent correct and makes large profits. The model’s optimizer adjusts the input signal variables to capture as much of that profit it can. In between the input and output processing elements are a number of “hidden-node” processes. Because of the complexity and interactions between the hidden nodes of a neural network, it is difficult to apply analytical techniques to understand how a decision is reached directly from the original inputs. So, we have to trust the output of the network blindly and have confidence in our design approaches.


We: Let’s go into the model a little more. When you talk about using standard technical indicators, how are you using them? How are look-back periods determined?

BR: We train them over 12 months. That model shows how well they’re doing, then we feed them into a genetic server, which takes the best inputs and optimizes the performance by figuring out what percentage of that signal should be used to work best over the past three months. We train them over 12 months in the neural engine and look back over three months in the genetic server to see how well they’re doing.


We: How long did you spend developing and testing your model before taking it live?

BR: A little over two years. There wasn’t a neural engine at the time that could handle that. What we developed was a context-memory neural engine.


We: Let’s look at some of your trade statistics, then. What’s your percentage of winning trades? BR: Over the past nine months the winning percentage has been 50.17 percent. CT: How about average or median profit vs. loss?

BR: The median profit was 1.478 times the median loss.


We: How long does a trade typically last?

BR: Two-and-a-half days on average.


We: How many different trades will you have on at a time?

BR: We have a money-management system that determines which pairs are doing well, and it turns off the ones that aren’t doing well. At any given time there are generally five or six pairs trading live and four or five that are turned off.


We: What’s the filtering criteria — that is, what does “doing well” in this context mean?

BR: When they aren’t winning as much as they are losing, we turn them off.


We: Is this a kind of equity curve management approach?

BR: Exactly.


We: Are you always trading your model or systems in the background on a simulated basis so you know when to turn certain currency pairs on or off?

BR: Yes.


We: How would you characterize the typical trade signal? Is it a breakout or trend-following trade, or an exhaustion or reversal type of trade? Or is there a mixture?

BR: In general, they tend be trend following. Just where on the trend it executes seems to be highly variable, though. The model takes positions in and out of the market. We don’t have profit limits. When we place a trade we also place a 50-pip stop, generally.


We: Regardless of the currency pair?

BR: Yes. We limit investor risk to 2.5 percent of his money, so it’s a constant risk model. It’s anywhere from 50 pips to 63 pips, depending on the model.


We: What kind of price data do you use in your trading model?

BR: Hourly data. To us, short term is hourly and long term is weekly.


We: Can you pick out a recent trade or two?

BR: A long trade in the NZD/USD, which was entered on March 15 at 11 p.m., is up 69 pips (as of March 20). It’s still on. We have another position on in EUR/CHF that started on March 16 at 8 a.m. and is up 85 pips.


We: Have you ever applied your modeling or trading techniques on other markets?

BR: Yes. We’re looking at the stock market now.


We: You had some huge monthly returns early on, followed by five up and five down months through September 2006. The last couple of years have been very tough in the forex market. How are you adapting?

BR: Well, the thing to be careful about when looking at our history, is that when hurricane Katrina came through, everybody got out of the U.S. dollar — that’s when we made that huge amount of money. We also have a new Fed Chairman (Ben Bernanke). He bounces things up and down, which is why November, December, and the beginning of January were a little bit different than we’d expected. But it looks like it’s turning around now.

CTWe: February seemed to be a good month or you (+17 percent). Did anything in particular contribute to that performance?

BR: The money management had turned the non-winning pairs off in January so we had a good-performing set of currency pairs in February.


We: How often do you adjust your trading model?

BR: We retrain it at the end of every month, over the weekend.


We: After a year or more, what have you learned about the markets and your model? Is it different from what you envisioned?

BR: We had to add what we call an “event processor.” We track nine different events — such as the first Fridays of the month (when many important monthly economic reports are often released). During the hour an event is supposed to happen, if a pair moves more than 30 pips in less than seven minutes, we’ll get out of the position. We update it every month.


We: Was that the result of taking big losses because of certain events?

BR: Yes, like on first Fridays — we learned that (laughs). That’s what some of the big drops last year were from. Forex is a market without any of the artificial market maker limitations, like short trading restrictions or the ability to drop through a stop position without executing. Banks have been in this market for centuries and now invest twice as much in the forex market as they do in the equities market. This is because they have developed effective training techniques of an intermediate time sample approach. Forex is nice in that it may be trading effectively on many different time samples.