Why was 2005 such a poor year for long-term trend followers yet many of the best performing programs define themselves as trend followers? Perhaps some programs had an added alpha boost. Those analysts who like to claim trend following doesn’t work are sure to be prevalent these days given the performance of commodity trading advisors throughout the last two years, but at the risk of offending certain sectors, it is clear that trend following works — though it will always experience years like the last two. A better question perhaps, given the variety of strategies that can loosely fall under the trend following label, is whether the term “trend following” is too broad to tell us anything regarding market performance.
After all, a trend is a pattern of sorts, but a trend is something that to a great extent reveals itself whereas pattern recognition signals are more complex.
Francois Bonnin, principal of John Locke Investments, uses various pattern recognition systems to help smooth out the return stream of the overall program, which includes a trend following approach. “Pattern recognition will capture very different motions in the market than those that are captured by trend following. Pattern recognition essentially can capture atypical moves rather than trending moves. An atypical move could be something that has nothing to do with the trend,” Bonnin says. He explains, “When you look at trends you look at something that will evolve and hopefully last, a pattern is something more spot like.”
Emil van Essen, president of Van Kar Trading Corp., says most pattern traders are looking for a certain combination of bars on a chart. “Pattern recognition, in its simplest state, is about defining what your pattern is, identifying it in history, finding out what the result was in history...when it occurred did it go up or did it go down and how reliably did it go up or down? And if they can say it went up with a lot of reliability then they have a pattern they can trade,” van Essen says.
A pattern can signal a trend, a reversal, a reversion to the mean or a breakout. It can be based on a certain configuration of bars on a chart or a spike in volatility. The different market movements that patterns can signal indicates these patterns are not simply trends of varying time frames. While many pattern systems are based on a shorter time frame than trend following systems, it would be a mistake to define them by time frame, says Ken Perseghin, head trader for Harbor Street Capital Management. “You can see the difference when you’re trading it and when you see the positions because the short-term trend followers have probably half of their positions opposite of us.”
OPTIMIZING VS. CURVE FITTING
As technicians move from simple to more complex trends using artificial intelligence and neural networks there is a danger they will find precisely what they are looking for.
“The trick is when you are doing pattern recognition, the more things you throw into the final, the more you are [overoptimizing], so what you want to do is [ask]: ‘Does my idea work in general?’” van Essen says. “The more things you add, of course, you will find something that will work. How do you know you have found something that works or if you curve-fitted? That is a difficult thing. How I would know, is that I would close in on that territory slowly and make sure that each refinement made in itself was profitable.”
Bonnin likes to keep his systems simple and numerous to avoid the curve fitting trap. “We have a precise approach to what we do. We don’t overfit any of our systems. Our systems have to be very simple, we have to understand what is the alpha driver, is this way of capturing alpha going to be present many times?”
Though to some extent pattern recognition is an attempt to find something in the market that is not readily apparent. Bonnin says creating arbitrary rules that appear to optimize a strategy is a sign of fitting and will eventually fail. If you do not know the why of a rule, it is better not to use it. “I have seen some systems that, backtested, are so good; but looking into the rules like ‘don’t trade Thursdays’ or ‘don’t go long if the RSI is below 30,’ and when you remove each rule you realize that two losing trades have been removed,” Bonnin says.
“Out-of-sample” testing is the best cure for curve fitting, van Essen says. Once a rule is in place it must prove its worth on trading data not used to create the rule. Van Essen says a common mistake is to use older data for the out - of-sample testing and newer data to refine the system. He says it is best to use more recent data for the out-of-sample set.
ALTERNATE VIEWS
Edward W. Dreiss, one of the founding fathers of technical trading systems, has developed systems based on the fractal wave algorithm. While his system methodology can be termed pattern recognition, Dreiss chafes at the idea of separating the methodology from trend following. “I am an exceptionally long-term trend follower,” Dreiss says. Dreiss refers to the search for highly technical patterns as, “like seeing horses in clouds,” and says, “Some systems may not work because of the reasons people think they work.”
While patterns can be the most technical of technical systems, they also can be applied to specific markets. It may seem counterintuitive to have a discretionary overlay to a pattern recognition system, but Switzerland-based Stewart Capital Management has one. Stewart principal Elaine Knuth says the problem with patterns is they can be similar but they are never the same. “They try and capture absolutes but there are no absolutes in trading. What I found in back testing is many of the classical patterns (head and shoulders, pennants, flags) are not valid. It comes out to 50/50,” Knuth says.
She also will read her proprietary patterns differently in different markets. She says the notion that for a system to be good it has to work in all markets is simply wrong. “Patterns are more or less valid in different markets. A V-shaped reversal in soybeans tells us a different story than in live cattle,” Knuth says.
ENHANCING RETURNS?
Modern portfolio theory tells us the more sources of non-correlated alpha a portfolio has the better overall risk adjusted returns it will produce. Several managers have worked that theory into their programs and incorporate multiple systems producing returns in different market environments. “If you can find sources of alpha, which are uncorrelated to each other, you should put them in a broad portfolio to have a better risk adjusted portfolio,” says Antoine Josserand, director of alternative investments for AXA Investment Managers. “We believe in diversifying sources of alpha within the broad portfolio so by using different type of models and different types of signals you have a better constructed portfolio in terms of optimizing your risk /return profile,” Josserand says.
Reversal patterns initiated in markets in the midst of long-term trends allow managers to offset losses during corrections and allow the overall program more tolerance to withstand corrections and profit from resumed trends instead of being stopped out.
This occurred in crude oil and Japanese yen in 2005. “If you look at oil last year, it was clearly in an uptrend for the first eight months and then a downtrend and then an uptrend. If you have a fairly long-term trend following model you are going to be whipsawed by this type of market move,” Josserand says. “Oil went from $40 to $70, and from $70 to $55 and then $55 to $65. If you trade long-term you are going to get long, then stop loss; get short, stop loss; get long, stop loss.” (See “Crude reversal,”)
Mangers who did well in 2005 made money on key reversals in crude oil and other markets as well as the broader trends in those markets. In some cases those reversal patterns simply offset the losses to their trending elements but, after all, that is the point. And by avoiding those whipsaws, or surviving them, they can stay in long-term trends.
“Pattern recognition signals can be active opposite a trend. It will have a shorter [holding period] than the trend [the market] is in because it is probably [finding] a consolidation of the trend. It can catch reversals; it can catch the beginning of trends earlier than the trend-following system when there is a nontrending market. It essentially can act and make money without the necessity of a trend and be there before a trend starts,” Bonnin says.
He notes trend following strategies were incapable of avoiding sharp reversals in oil and other markets in 2005 because the original trend was so strong it required a strong move to signal a reversal. The patterns allowed him to neutralize the effect of the correction. “Stop losses are faster in the patterns because we know what we are expecting in the move and if it doesn’t happen, we kill it,” Bonnin says.
Adding pattern recognition systems into an overall portfolio allows a manager to let the trend trades run a little and not keep trailing stops so close that they will be hit before a trend resumes. “It is very difficult to figure out when the trend stops so we have to leave a little bit of space for the market to breathe. The patterns look at these breathing moments because breathing moments are very often along the conditions of a pattern,” Bonnin says.
Harbor’s Perseghin says his pattern systems have outperformed his trend following systems throughout the last two years. “The markets have been very choppy and I have even seen the markets expand their ranges [while] still being choppy.” That is a particularly difficult environment for trend followers but one that can be exploited by patterns.
Bonnin agrees. “I don’t see pattern recognition or trend following as complimentary to anything, both approaches derive alpha. To build the best portfolio is to have several alpha diggers that will capture alpha in certain circumstances that are repetitive such as trend following or certain patterns. They both dig alpha at different moments, if you blend them together you will create a portfolio that will be capable of adapting to the market much better.”
The test for many of these strategies that perform well in tougher trending environments will be how much they hurt the overall strategy during good trend following environments. Perseghin says their short-term pattern recognition strategies tested well in 2002, which was a great year for trend following.
STAND ALONE
While Perseghin says their pattern recognition models help enhance their overall returns, he doesn’t just see them as a filter or as a compliment to a trend following strategy. He says they can stand on their own and plans to offer a fund using them exclusively. “We are looking to use them exclusively and also to compliment.”
Bonnin however, likes to have his models work together. “We are not trading one magical idea, we are piling up some very clean and robust ideas that taken individually are not so fantastic but because we have several that each perform a specific task, when added together they slowly capture more and more money from the market.”
While John Locke uses three separate pattern recognition strategies: mean, reversion, volatility breakout and reversal, Bonnin is not considering eliminating his trend following approach. “At the moment the strategy is not good, but in the end trend following will be a superstar again.” Using patterns during difficult trending periods allows managers to earn solid returns even when their primary strategy is not a star.
