More Random Bits of Data: Asset Class Correlation
These are more bits of data excreted by my research into building a tactical asset allocation model. Unfamiliar with TAA? Click for a primer.
In this post I show the correlation between eight asset classes, how that correlation has changed over time, and why that might be important to a TAA model.
The first table shows monthly correlation between all combinations of asset classes since 1970 (data permitting). But correlation is notoriously hard to pin down because it’s constantly in flux. The next table looks at just the last 5-years (ending 08/2010).
In this shorter view we see much higher correlation in most pairs, especially between anything related to stocks/real-estate, and between commodities and stocks.
I think reasons for changing correlations are three-fold:
1. Growing Interrelatedness: Some pairs (ex. anything stock-related) have seen a constant and steady increase in correlation over time. Chalk that up to globalization.
2. Oscillating: Some pairs (ex. stocks vs Treasuries) have always oscillated high and low (and high again) through different market regimes.
3. Market Shocks: Some pairs become temporarily interrelated (ex. oil and stocks) due to market shocks and I think can be expected to return to historical “norms”.
How does all of this relate to our TAA model?
We’re deciding what asset classes to hold in any given month based on trend-following/momentum, but we’re deciding how much to invest in each of those assets partly based on how much each asset is expected to contribute to portfolio volatility.
Part of that contribution is the volatility of the asset itself (i.e. more volatile assets are allocated a smaller % of the portfolio), but part is also how well each asset is correlated to the others (i.e. less correlated assets are allocated a larger % of the portfolio because they cancel out some portfolio volatility).
Side note: the correlation figures above are meant to be instructive, but they’re insufficient. What we really need to understand is correlation during market shocks (ex. the 2007-09 bear market) because, as the hyperbolic saying goes, during crises “all correlation goes to one”.
The point of all of that is to say that understanding how all of these asset classes actually relate to one another TODAY is vital to a “smart” TAA model, but as shown above, that’s difficult because correlation is a constantly moving target.
[Edit: click for a summary of all posts in this series on TAA]
Happy Trading,
ms
. . . . .
To stay up to date with what’s happening at the MarketSci Blog, we recommend subscribing to our RSS Feed or Email Feed.
Filed under: Tactical Asset Allocation | 5 Comments
[Side note: the correlation figures above are meant to be instructive, but they’re insufficient. What we really need to understand is correlation during market shocks (ex. the 2007-09 bear market) because, as the hyperbolic saying goes, during crises “all correlation goes to one”.]
That should be “all correlation goes to one or minus one”. Some safe-haven assets (on a global scale) went to minus one during the crisis: think USD and US treasuries.
Could be interesting if your TAA scheme could incorporate this “flight to safety” instead of just going to cash…
Kind regards,
-Mark-
RE to Mark: That’s why I called the statement “hyperbolic”…I agree that it’s not true that all correlation goes to one. But good point.
In theory, if any asset is existing positive trend/momentum, it should be picked up by the model. So for instance during the 2007-09 crises, the strategy would have rotated into assets that (implicitly) had low correlation to equities (otherwise they wouldn’t be in an uptrend at that moment).
I’m not considering USD at the moment (but it’s on the short list of asset classes to look at).
michael
There is a matlab package for dynamic conditional correlations. Basically think of Garch, but instead of time-varying volatility you have time varying correlation. I have also tried using regime-switching for the correlations, but that seemed to take a very long time with only five assets.
It would be a wee-bit tricky to implement for most. I would ideally use Meucci’s Entropy Pooling framework which allows you to take views on correlations, but admittedly I’ve only taken more simple views. I assume that after you simulate out some returns with DCC you would need to fit them with a copula to use Meucci’s method, but I’m not sure entirely.
MS, don’t forget the importance of interest rates to the correlations between assets, such as globally-traded commodities (gold and oil) and stocks. Because USD interest rates are low, the USD is a borrowed currency in the “carry trade,” which means that USD borrowing funds purchases in the risky assets of stocks, so that now, unlike historically, stocks are inversely related to the USD, just as oil and gold pretty much always are. I think that is the main reason for the high recent correlation between oil/gold and stocks. you will find the correlation increased right as our interest rates dropped. if you want to isolate correlation unrelated to USD correlation, you have to price these assets outside of USD movement.