I’m trying to organise the major pairs into groups based on their correlation with each other. So when I backtest, I’ll only test one pair from each group. This will help avoid duplicating highly-correlated trades and distorting my backtest results.
These are the groups I came up with:
European Group: EURUSD, GBPUSD, USDCHF
Japanese Group: USDJPY
Commodities Group I: AUDUSD, NZDUSD, AUDJPY
Commodities Group II: USDCAD
Cross Pair Group: EURJPY, GBPJPY
Any thoughts on these groups? Do you think there is enough correlation to justify grouping these pairs? And is only backtesting one pair from each group a good idea?
It’s really pretty simple. Any pairs which share a common currency (EUR/USD, USD/JPY, USD/CAD in the case of USD for example) will tend to be well correlated. Pairs which don’t share a common currency will tend to be less so (though some currencies - like EUR and CHF right now - are closely linked at different points in time)
The last time I checked, the 1 year correlation between EUR/USD and a few other pairs containing USD in them are as follows:
EUR/USD and AUD/USD: 0.08
EUR/USD and NZD/USD: 0.11
EUR/USD and USD/CAD: -0.04
EUR/USD and USD/JPY: -0.19
As we know, anything between -0.2 and 0.2 implies weak to negligible correlation. So I beg your pardon, your generalization is really pretty oversimplified.
In fact, there is more value in KK’s original classification. From the above correlation figures, we can see that the commodity currencies (AUD, NZD, CAD) and USD are highly correlated, and not simply because there is USD in them.
My initial response is that your grouping would probably depend very much on your intended backtest duration. Correlations fluctuate quite drastically over time, so whether you are backtesting for 10, 5 or 2 years makes a significant difference I would think.
Secondly, based on my own experience, more often than not, currency pairs that are supposedly highly correlated tend to give quite different results depending on the precise strategy. Hence, I’ll probably not take the shortcut and instead backtest all interesting pairs that might potentially be suitable for the strategy in question.