# How to extract informative value from correlations of assets? Subadditivity of correlation calculation an issue

I was reading Nassim Taleb's Paper: Fooled by Correlation and found it very informative. I had always struggled with finding value in correlation in Finance, especially seeing a lot of bad applications in the workplace, however this paper got me thinking.

Would splitting up correlation into subsections as shown here have informative value for model building?

Say I wanted to model the % price change (or the log % price change) for a pair of assets, would it be better to look at the correlations for the downwards movements and upwards movements separately?

Are there any risks to doing so for lagged correlations between the variables?

A more general question as well, how do we prove causation in asset price relationships with different features in our model?

Also, do there exist any Python libraries that calculate these subsets of the overall correlation?

• Hi: I only read the beginning of the article ( section 1 ) but I think the idea regarding subadditivity is that, under the bivariate normality assumption, the correlation number, $\rho$ is very dependent on where one of the numbers ( X or Y ) is in the distribution percentile wise.. That makes sense, but when correlation is calculated in practice, I generally don't see where the normality assumption comes into play so I'm not clear on the use of what is shown in Section 1. Jun 5 at 22:51
• Well, for these kind of applications you'll need econometric models like the ECM. Jun 6 at 8:23
• And please take the papers of Nassim Taleb with a huge grain of salt; his work can most of the time not be tested or can be falsified. On twitter he was explaining how to solve the P*NP problem and that he had done so but, as a real showman, he deleted the tweets after some hard questions from theorists. Jun 6 at 8:31
• @simsalabim thats hilarious, did not know that. Jun 8 at 4:25