Places I’ve Been
I have just been on a visit to colleagues and clients in Asia, and it was an almost exclusively agreeable experience. We at Macrobond are indeed making an effort to add to our database more EME-data in general and Asian data in particular (yes, a lot of it is for China and India), but I was actually quite surprised to hear almost no direct complaints about data coverage, etc. Now, would any of our readers disagree – or have other input, I am all ears. This is important to us.
What I did not have the time for (this time) was to meet up with our clients Down Under, why I thought I owe our mates there at least a blog post. This is it. For the rest of you, I hope it provides some inspiration about how to delve into the workings of your local stock indices and (one way to) explore what’s going on.
On a Walkabout in the ASX200
Quite a few analysts have used the PCA (principal components analysis) on fixed income markets, and some have also tried to use it on stock markets. And I kind of like the idea of using PCA on the stock market constituents as it first and foremost serves to “reduce dimensionality” in data. With some exploration (and imagination), we should hopefully be able to explain a large part of how and why stock indices (in this example the ASX200) move. I chose a cut-off point around 1998, to preserve at least 50 companies (index constituents) in the calculations. Ideally, we would prefer very long time series, but the problem here (besides survivor bias) is that the constituents tend to merge, split, list and delist or, plain and simply, go bankrupt. Well, anyhow, this is what the result looks like:
These four components explain almost 60% of the variance in the ASX200, but when experimenting with longer (or shorter) time spans, i.e., reducing (or increasing) the number of variables/constituents, I can easily push that up to around 70%. In the end, the choice of variables and components is of course each modeler’s decision to make. In my analysis, I have identified four components, and while I might have been able to explain additional components (PC5, PC6 etc), time is a constraint, so I settled for the four first components.
When performing PCA on a stock indices, the first component is (almost always) expected to relate to the market risk (movements in the headline index). And sure enough, the correlation coefficient between PC1 and ASX200 is above 0.9, which is visible in both the graph above (look at the red and black lines) and in the graph below:
What drives the headline stock indices? – Well, the correlations are not as straight-forward as one would like (because the constituents might not be a perfect representation of the economy), but clearly economic growth “should” be a decisive factor. And in this analysis, I have used GDP and Industrial Production, but I’m sure PMI-data or anything else would work just fine. When you study the matrix in the document (download and open the chart), it is also quite obvious that cyclical and financial companies are more strongly influenced by the first factor while utilities etc. are less responsive to the PC1.
Extracting and exploring the second factor, I expected strong covariation with inflation measures, which consistently yielded high correlation coefficients – especially for producer price measures – but other measures could also come into play (inflation expectations, break-even inflation etc.). Feel free to explore any connections yourselves!
The third factor was surprisingly hard to pin-down, but after many essais I believe it is related to commodities and, in particular, those related to industrial raw materials and foodstuffs. If you happen to come across something that works better, I hope you let me know.
As difficult as it was to connect the third component to a clearly defined macro variable, the fourth component was something of a breeze. Obviously, the Aussie dollar is very strongly correlated to PC4, and it seems to be more a question of having the right currency measures, rather than interchanging the currency for another variable (I tried interest rates, btw).
Taking it Home
Today’s exercise was hopefully somewhat illuminating. With just a few simple macro variables, we are able to explain most, up to 70%, of the variance of the ASX200, which is rather impressive. Our results suggest that “only” 30-40% of the risk is truly idiosyncratic. There also seems to exist a nice lead/lag structure that you might be able to explore. For some of the components it would be interesting to continue our exploration to try to find a macro variable that ties in harder to the first and third components, for example. A lot more information could probably be extracted as well when studying the matrix of the eigenvectors more closely (what kind of companies seem to correlate the strongest with the different components, e.g.). Likewise, it would be interesting to add one or two additional components to the analysis (to explain more of the total variance), and I’m guessing domestic costs, like wages and interest rates, would come into play as we delve deeper into the principal components. That said, at some point, the “residual variance” is probably particular to some individual constituent(s), so it is probably of little use to wander too far away from home.
 Or, to be truthful, of the 50-odd companies we used in the sample.