One thing I’m greatly interested in at the moment is branching out my fleet of strategies. Ideally I would like to have multiple uncorrelated strategies trading at the same time such that when one is in draw-down, the other is making money. Not only does this apply to multiple different strategies trading simultaneously, but it could even apply to the same strategy trading across multiple instruments.
Recently I have been playing with some Bollinger based reversion strategies. One of these I have optimized for several currency pairs, and wish to run it on multiple instruments using different configurations. While they all make money overall, they could be correlated and many of them could draw-down simultaneously. So how do we measure/check this?
The best way to check this of course would be to back-test it. However, Currently the back-testing tools in FXCM Trading Station only support one strategy at a time. While they support multiple instrument’s price data feed, I would have to modify my strategy to subscribe to the multiple instruments and effectively trade on them all at the same time. Not only does this complicate the strategy’s implementation, it also restricts the amount of data I can test for historically, i.e., the more instruments you subscribe to, the shorter the period you can test for.
So, what else can I do?
While researching this topic, I came across a post on Mechanical Forex that suggests an interesting idea of examining the correlations in the monthly returns of each instance. If you do this, you may be able to identify which instances make money or lose money at the same time. I thought I would test this out.
Using the idea on the above mentioned post, let’s try to find the correlations on the returns. Since the reversion strategy I am using trades very frequently, examining monthly returns is not useful here. What I decided to do is, modify the strategy to log the equity at each hour over a 1 year period.
I then took this data and calculated the correlation between the hourly equity values for each instance.
It’s great that it identified that USDCHF is not really correlated well with USDJPY, but the rest of the instrument’s correlations were too close to really gain much value. I realized that since all the strategies make money in a fairly linear fashion, if we try to find the single correlation value of over the near 6500 data points, they probably will be highly correlated. It does not make sense to try to summarize the correlation in a single value.
Therefore, I decided to try to visualize this data using a site I recently discovered called DataCopia. DataCopia allows me to upload a bunch of data in excel format, and it automatically generates a range of charts based on this data. Sometimes it can be good to identify patterns by visualizing the data in ways you could not have imagined.
Plotting these values in DataCopia we get the following line chart.
Normally DataCopia produces many charts, and sometimes it can break data down and find sub-correlations and things like that. Unfortunately in this case it only produced a single chart, I think because there was so much data. Regardless, by visualizing this data we can identify points in the equity curve where multiple instruments have taken a hit at the same time and the equity has dropped lower, or they have been on winning streaks and the equity has trended higher.
This kind of information can be useful in determining which strategy instances should be run together or not.