Sunday, February 14, 2010

Stop it with the data-mining, Mr. Rosenberg

I have a lot of respect for David Rosenberg, the chief economist over at Gluskin Sheff (previously with Merrill Lynch). However, I have a lot of trouble reading his daily commentaries because: 1) they're a tad too long, 2) they often repeat, and 3) Rosenberg much too often cherry picks data to fit and prove his preconceived opinion (opinion to which I agree for the most part - the point being I much rather like people forming an opinion out of data and logic rather than the other way around).

As an exemple, Rosenberg recently highlighted the extremely high (0.95) correlation between the S&P 500 and the Copper/Gold ratio over the past three years. Unfortunately, the reason for this is more the fact that the correlation between stocks and both the metals has been very high over that period. In econometrics, this is known as multicollinearity (a violation of regression assumptions), and is typically detected by a high correlation which carries low significance, reflecting inflated standard errors.

Let's look at the data since 1997: the correlation between stocks and the copper/gold ratio drops to 0.55, while the correlation between stocks and copper is a lower 0.37 (the one between stocks and gold is about zero). 0.37 is still statistically significant so we haven't got rid of the multicollinearity problem, but at least now we can overlook it. Conclusion: Copper/gold has a somewhat significant and relatively high correlation with equities.

Mr. Rosenberg, you see, your point was valid. Why did you have to torture the data to "prove" it?

No comments:

Post a Comment