Have a plan
The first thing is to have a plan. If you built your system skillfully, it may respond to the market reasonably most of the time. But it can’t respond to correlation spikes reasonably because no matter how thoughtful you are and how large your sample size is, it is impossible or almost impossible to get a meaningful sample; correlation spikes are rare. So, it probably isn’t in the system. The history of the market is, to some extent, a history of non-recurring events. How, exactly, are you going to plan for that? If you are extremely alert and know what an unexpected event looks like when you see one, you may need to do nothing more than constantly scan the horizon for danger, condescending to act once every decade or so. If the correlation between wheat and corn goes above 0.9, what are you going to do? Are you going to close all of your corn positions and keep a little bit of your wheat? Close out all of your positions? Close some of both and see what happens? You need to know what you are going to do before the market hits the fan.
Not incidentally, deciding you are never, ever going to override your system is a plan and for some fund managers it may be the right plan. But that plan means that every now and then you are going to lose on just about every trade you have. The rest of your money management system has to be built so it can withstand those kinds of shocks. In which case, your best bet may be to reduce your leverage and take your hits like a man. Correlations shift and like all elements of a trading plan, it is good to design a strategy for it in the calm during the system building process rather than in the heat of a fast market.
Watch the markets
When correlations rise, lower your commitment. With the right software, watching the market is a lot easier than it sounds. Unfortunately, all of the commercial software I know does it wrong. Here’s the problem: most correlation measures measure the relationship between data pairs. Which means that if you are following N markets you have to follow N(N-1)/2 pair wise correlations. For 10 markets this means 10(10-1)/2 = 45 pair wise correlations. This can be done by watching half of a matrix (the top left to diagonal right values are all ones and the bottom of the matrix is a mirror image of the top) but it can’t be done well. Unless the matrix is tiny, it is impossible to keep everything in mind. But the world comes in natural groups (interest rates, stock market indexes, grains, metals, oils, etc). By using a measure such as Kendall’s W, you can measure the correlation of a group of markets. In which case, the number of correlations the trader has to watch may well be in the middle single digits. Kendall’s W, also know as Kendall’s coefficient of concordance, can be found in most books on non-parametric statistics. It’s not that hard to program.
Find non-correlated trades
If possible, look for trades that are not correlated. Obviously this goes beyond our normal markets that have become correlated. Bizarrely, trading profits are not always dependent on changes in the price of the underlying investment. Bertrand Savatier of Numbers Alternative Management, a French CTA, trades calendar spreads, among other things. Mr. Savatier says “calendar spreads aren’t correlated with anything, including the underlying investment.” Savatier also suggests using fundamental filters such as seasonal approaches because they necessarily differ for each investment.
If you use software to allocate your trades, don’t overfit. Money management software bases its allocations on estimates of return, risk and correlation. Change a couple of those estimates even a little bit and your optimal money allocations can change in radical and unobvious ways. Worse, professional money management software is usually based on systems that are almost impossible not to overfit. So, for N markets, most such systems demand 2N + N(N-1)/2 estimates. For 10 markets a system based on quadratic programming, for example, demands 65 estimates. How much do you want to bet that a few of those estimates aren’t going to be a bit squirrelly?
Make sure you understand how your markets correlate. Start by finding out what the baseline correlations are. The futures markets for wheat and corn, for example, are more than a century old. Are wheat/corn price correlations stable or do they cycle over years or decades? When correlations depart from their historical norms how long does it take them to go back to their norms? Hilary Till, who runs a private trading firm, argues, in effect, that the recent history is not necessarily relevant. In Till’s opinion, there are types of markets. Forecasting price behavior and risk properly depends upon finding similar past markets.
When we last spoke to “Paul,” we hadn’t listed all these ways to avoid the risk of rising correlations. “Paul” looked over at me and said, “Saw this James Caan movie a couple years back. Caan says, ‘If you can afford to lose, you’re not gambling. I’m a gambler.’ You see that movie? Anyway, I’ve done my share of gambling over the years. I think I would rather just keep some of
Fred Gehm is a consultant to CTAs and the publisher of FredGehm.com. He is the author of “Quantitative Trading and Money Management” and the forthcoming “Trust is not an Option: Evaluating and Selecting Investment Managers.”