Correlations and trading

Correlation not causation. Ok, we’ve heard that thousands of time before. The difficulty is that we cannot simply ignore correlation if we are trading financial markets. The relationships between different assets and other data is often crucial to forecasting the market. If we were forced to ignore it, you’d end up ignoring so much data and instead be forced to look only at price data. Even with price data driven strategies such as trend following we are relying on some element of correlation (in this case autocorrelation within a time series).


There are many examples where we might use correlations in markets to come to a view. If the dollar has sold off significantly against most currencies expect one, is there a catch up trade there? If we can forecast one data point such as GDP or understand how central bank policy could change, how does that impact rates and FX more broadly? It’s all about thinking about the chain of events, which are often linked, and how they impact each other. The crucial point is that we need to somehow ignore what might be spurious correlations. What can be equally puzzling, is when we are able to make an accurate forecast about one event, such as a data release, which has an intuitive relationship with the market, but the market simply ignores it. 


Earlier this week we saw an example when correlations suddenly moved in reverse to expectations in USD/JPY, which steamed higher, driven by fears around the Coronavirus on Japan, despite a general risk off tone to markets and a fall in yields. Traditionally, JPY tends to appreciate during risk off (ie. USD/JPY falls), on a safe haven effect, driven by several factors. These include the fact that JPY tends to be a funding currency and also because of spot being driven by longer dated USD/JPY vol instruments. Even for example during the Japanese earthquake, JPY appreciated significantly, and USD/JPY reached record lows.


So what do you do if the correlations go out of whack, which your trading strategy or approach uses? The first thing to do is to understand how the distribution of returns looks like. Is it consistent with other historical periods for your backtest? If it is totally different, and you can’t explain what has happened, maybe your model is “broken”. For higher frequency strategies with a high Sharpe, it is often easier to identify this point, what was a straight line of returns deviates suddenly. Is the factor you are modelling likely to have disappeared? Have other market participants had reason to change what they do? Or was there a one off event which caused this deviation, which is unlikely to repeat, and hence could the strategy comes back? From experience the time when you are desperate to turn off a model is usually the point when it starts to perform again!


For longer term strategies such as trend and carry, is it more difficult to identify these “breaking” points, and often they can be subject to extended periods of underperformance. Again, you need to try to understand whether the factor you are modelling still has relevance. For trend, which is basically based on herd like behaviour in markets, I think it would be difficult to say this has gone away. You only have to look at Tesla stock recently to see an example! For developed markets FX carry, it did underperform for a while, in part given there isn’t much of a rate differential in markets, but came back strongly in 2019. It seems to make sense when there is a larger rate differential in markets carry should perform well, and when vol is also low. Admittedly, these are only to examples of anecdotal evidence, but they nevertheless illustrate some of the quandaries you face when correlations shift. There are several well known examples of strategies which suddenly begun to underperform, as many market participants were forced to liquidate, such as the quant quake of 2007, which impacted statistical arbitrage strategies and the quant quake of 2019, that caused suffering for equity momentum strategies.


So yes, correlation is not causation, yet these relationships they can tell you a lot about a market. The difficulty with correlations is they are not always stable, even if we can eliminate the spurious ones. Correlations can be your friend, but always be wary!