One of the arguments against international diversification is that correlations among markets increase during periods of increased volatility (ex: crash of October 1987, most of the world’s markets were down sharply). The explanation to refute that argument in Schweser is that “correlations increase as volatility increases due simply to the statistical properties of the correlation measure.” What does that mean exactly? I know that Correlation = Covariance/Variance Does that mean the numerator increases more than the denominator?
AKA: “Correlation Breakdown” you have to think about the formula for covariance. as markets become increasingly volatile, the markets begin to move together. this is a legitimate concern in an increasingly global market. Think about what happened on MLK Day. All the markets around the world sank together–there was little benefit to being diversified because the correlations between and among markets moved closer to 1 during a period of upheaval/volatility.
But you are making the case against international diversification, no? What I don’t understand is the argument that says that correlation does NOT increase during major recessions or depressions. I believe that correlations do increase. To use your example, on MLK day, all markets were sharply down.
"The explanation to refute that argument in Schweser is that “correlations increase as volatility increases due simply to the statistical properties of the correlation measure.” That is just nonsense and I can’t imagine what they might mean from that. Correlations among markets absolutely increase when it hits the fan.
That’s what I thought too Joey. Here’s another excerpt: “The problem with estimating correlation during periods of rising volatility is that the correlation will be biased upwards when in fact it has not changed. If we take a sample of data and split the sample in half based on the absolute value of returns, the high-return half (with an accompanying higher volatility) will have a higher measured correlation, whereas the low-return half will have a low measured correlation. So although the true correlation might be about 0.40, the higher volatility of the first sample exhibits higher correlation. This phenomenon is due to the econometrics of correlation measure. Higher volatility in returns equates with higher measured correlations. Academic research has found that the previously reported increases in correlation during volatile stock markets were, instead, manifestations of the higher volatility and not increases in the true correlation.”
Ok I read some of those papers (e.g., http://www.bis.org/publ/confer08k.pdf). Idiotic and repeatedly making statements just like the one above. So the problem with that glib explanation is that you have taken a bivariate distribution and split it into two different bivariate distributions. For example, if the initial distribution was bivariate normal, the two distributions you get when you split the sample aren’t bivariate normals with different volatilities, they are not even bivariate normals. In fact one of those distribution has zero probabilty mass anywhere around 0 which is about as unnormal as you get. And then they say the difference in these samples is their volatility?! I need a vacation from the world.
Not sure I understand the math above, but it seems possible that if large movements are correlated (i.e. when people panic, they panic about pretty much everything nonselectively, and when they think they’ve overreacted, they over compensate with everything, but when they are not panicking, things are less correlated), then the covariance and correlation might be disproportionately dominated by those large movements. However, the solution to this sounds like some kind of regime shifting model, where you have a “panic” correlation and a “non-panic” correlation, and not simply to assume that the correlation between assets long term is somehow biased upwards.
can you tell me what book/page Schweser this concept is in? thanks,
If a dataset suffers from heteroskedasticity, it should be ‘repaired’ before we can use it to come up with statistical figures based on it (something discussed in Level 2 quant section). Statistical figures are anyway measures which have limitations. Correlation is a measure that doesn’t factor in large changes in volatility, so isn’t relevant anymore if raw data before and including a crash, which is very probably hugely heteroskedastic, are used. So, we can argue that ‘correlation’ itself hasn’t changed during volatile periods. But the ‘usefulness’ of international diversification diminishes during volatile periods, and this has nothing to do with correlation per se. One example: your gf is hot, and you are pretty much sexually attracted to her. Say, her sister comes over who is twice as hot. Her sister being twice hot doesn’t decrease your wife’s “hotness”. But it decreases the utility of her hotness to you. Because you are now more attracted to her sister, and want her more.
That’s very silly
correlations increase because its the same guys which are investing here and there and all over, when these guys have to sell (ie redemptions etc) they will sell everywhere