# How do I graphically represent the evolution of a covariance matrix over time?

I am working with a set of covariance matrices evaluated at various points in time over some history. Each covariance matrix is $N\times N$ for $N$ financial time-series over $T$ periods. I would like to explore some of the properties of this matrix's evolution over time, particularly whether correlation as a whole is increasing or decreasing, and whether certain series become more or less correlated with the whole. I am looking for suggestions as to the kinds of analysis to perform on this data-set, and particularly graphical/pictorial analysis. Ideally, I would like to avoid having to look in depth into each series as $N$ is rather large.

Update

The following graphs were generated based on the accepted answer from @Quant-Guy. PC = principal component = eigenvector. The analysis was done on correlations rather than covariances in order to account for vastly different variances of the $N$ series.

-
Those charts look great - thanks for posting! This seems like an interesting tool for identifying regime change points. –  Quant Guy Aug 2 '11 at 21:29
Multivariate DCC is another way. There is support for this in the rmgarch package. –  Jase Dec 27 '12 at 10:12
For folks wanting a quick intro to Eignevectors, Eigenvalues and vector angles (used above): riskprep.com/all-tutorials/36-exam-22/… –  DeepSpace101 Jan 6 '13 at 16:27

I would consider a motion chart that plots the eigenvalues of the covariance matrix over time.

For a static view you can create a table: rows represent dates, and columns represent eigenvectors. The entries of the table represent changes in the angle of the eigenvector from the previous row. This will show how stable your covariance structure is.

You can also create a second table this time with eigenvalues as the columns sorted from high to low (and the corresponding values below for each date). This shows the variance described by each eigenvector so you can see whether correlation as whole is increasing or decreasing

Update: You can also measure the distance between the two covariance matrices via some distance measure metric such as Kullback-Leibler divergence, euclidean distance, Mahalanabois, etc.

-
This may be a bit abstract but here goes. Create a table: rows represent dates, and columns represent eigenvectors. The entries of the table represent changes in the angle of the eigenvector from the previous row. This will show how stable your covariance structure is. You can also create a second table this time with eigenvalues as the columns sorted from high to low (and the corresponding values below for each date). –  Quant Guy Aug 2 '11 at 19:31
That's brilliant! I just tried it and it actually works really well to demonstrate major shifts in the correlation structure. Can we do some cleanup and incorporate your comment into your answer? –  Tal Fishman Aug 2 '11 at 20:00
Cool! I updated the answer. I'm not sure if you are able to share/upload that output but it would be very intellectually interesting to look at! –  Quant Guy Aug 2 '11 at 20:09
Be mindful of the algorithm used to produce the eigen decomposition, since some methods (thinking eig() in Matlab) are arbitrary with regards to eigenvector sign, so the angle of rotation may not always be consistent. Another option to illustrate similar features is to plot the percentage of variation explained by a fixed number of principal components (sometimes called the absorption ratio) over time. –  michaelv2 Aug 3 '11 at 12:34
@michaelv2 you are absolutely right, and in this case I took min(x,180-x) for the graph above. I think percentage variation explained is what quant-guy had in mind initially, and I plotted that as well, it is also helpful. –  Tal Fishman Aug 3 '11 at 23:26

Here's an interesting possibility: correlation network analysis + motion chart.

Thanks to the hot research efforts in social network analysis (SNA), network analysis and graphics libraries such as R and Gephis are now easily accessible. I am well-versed in correlation analysis, and have a feeling that SNA can be effectively adapted for it. After all, the 'linear' correlation is just a special relationship among many others. Since SNA is hot area now. Many efforts are devoted and many resources are available. Can't help imagining that we can leverage it in finance.

Then the motion chart concept can further leverage the power of network representation. I would imagine that when a regime shifts toward high correlation, we should see the dynamics of 'clustering' effects, whereas low correlation regimes would show nodes scattered around. For example, if there is a sector-level event, we would see nodes in that sector start to cluster more tightly around few benchmark nodes. In a macro event, we would see all benchmarks or eigenvectors cluster.

In any cases, I think the motion network graph will be a much richer representation than the motion matrix, where the dynamics are usually represented in numbers, colors, or angles. If one still prefers to visualize 'covariance', then the node size will be a natural place for volatility (though I still prefer to separate correlation and variance visualization). The line width / color / distance can represent something else intuitively.

Last, I guess the reason evolving network graphs are not often used for financial correlations is that a lot of the traditional financial applications are static, in which case the advantages of network graphs over matrix representation are limited.

-
Interesting idea. Do you have any references of how it can be adapted to correlation analysis? –  Tal Fishman Aug 4 '11 at 20:54
you may be right that SNA could one day prove useful in finance, but I'm really not sure that this particular application (my question) is such a case. Your answer strikes me somewhat as falling under the saying, "to someone with only a hammer, everything looks like a nail." Nevertheless, I encourage you to post your own question asking whether SNA is useful in finance. –  Tal Fishman Aug 4 '11 at 22:00
I believe (more precisely, I 'guess') their underlying analysis is very similar (but please feel free to correct me if I am totally wrong). It should all start the cluster analysis, en.wikipedia.org/wiki/Cluster_analysis Then, apply certain network visualization algorithm en.wikipedia.org/wiki/Network_visualization Here is an attempt (not exactly the same way I suggest) I have seen: investuotojas.eu/2011/03/22/correlation-network –  楊祝昇 Aug 4 '11 at 22:10
I take back my earlier comments, I think this approach is interesting, too, and has its merits. –  Tal Fishman Aug 8 '11 at 4:23

I'd look at the evolution of a heat map based on the correlation structure (literally the lower triangle). I'd probably write a script in R or python that writes out the heat map per t to disk, then use a command line program like imagemagick to stitch images together into an animated gif, for example. I'm sure you could do it entirely in Processing too, and there you'd be able to mouse over pairs, etc. Perhaps seriation techniques could group sets ex post within the triangle.

-

Minimum spanning trees are another option, with edges between nodes based on either Euclidean distances of the matrix or another distance measure of your choosing. They can be more effective at illustrating the underlying structure of the matrix than some other methods (eg heatmaps, eigenvalue ratio plots), but this may not be practical if T is large.

You didn't specify what environment you're working in, but R has several packages (see vegan:spantree, ape:mst, igraph:minimum.spanning.tree, ade4:mstree and fAssets:assetsDendrogramPlot or fAssets:assetsCorEigenPlot) that support plotting both minimum spanning trees (with various layout types) and dendrograms. Gephi also produces plots that may be more aesthetically pleasing (and it's open source), although it does require that the data processing be done elsewhere and there is a bit of a learning curve.

-
I think you misunderstood my question. In my case, it is pretty clear to me what the relationships are between the variables, and in fact I have already plotted dendrograms to investigate this. However, it is not clear how I can examine changes in the inter-relationships over time. $T$ is about 600 in my case, so clearly drawing a minimum spanning tree and/or dendrogram at each point in time is not practical. –  Tal Fishman Aug 2 '11 at 18:07
Agreed. The only drawback to eigenvalue plots in general is that it's impossible to determine which variables correspond to which principal components, but if you only have a few matrices to evaluate then MSTs don't suffer from this problem. –  michaelv2 Aug 3 '11 at 14:19
This appoach could be use if we first make cluster analysis, and then make minimum spanning tree for cluster centres. Beyond that change in spanning tree structure could be view as evidence of some "topological" changes in market structure. –  Qbik Apr 26 '12 at 12:33

On conceptual level making cluster analysis or kmeans, for arbitrarily chosen k (in cluster analysis with hierarchical methods we simple would cut dendrogram into k pieces/subtrees) and then copmuting average correlations is much simpler then PCA.

But there are some problems with cluster analysis on correlation matrix of time series. If ,for each day, we have one correlation matrix of price changes with 15 minutes time resolution and then clusterize it into K pieces, then for each day we could have different sets of stocks in each cluster, we can fight or use it.

Fight it: if data are from the period of 6 months then we find clusters by merging daily time series of each stock into one time series and then clusterize and compute averages or some different measure. For Future data we can hold stocks memberships to clusters or merge new observations with previous set or make analysis in moving window.

Use it : compute total number of changes in membership of stocks - first day we have clusters A,B A={stock1,stock2} B={stock3,stock4,stock5} next day : A={stock1,stock3,stock4} B={stock2,stock5} so 3 stocks have changed its clusters. The more stocks have changed clusters the more market behaviour is unusual that day. This could be done easy with kmeans - we take means from day T, and attribute memberships for stocks T and T+1day, then count differences.

and distance matrix =(2*(1-corelationMatrix))^0.5

Some cool visualizations of correlation matrix by networks analysis : http://www.maths.tcd.ie/~coelhor/Palermo_Presentation_v1.0.pdf

-