Let's say that I have a universe of stocks from a certain sector. I want to compute the market portfolio of this sector. Beta is the covariance between each stock and the market. But how do you compute the market portfolio?

I read in several papers that a proxy for the market portfolio of a universe of securities is the portfolio with weights equal to the first principal component of the correlation matrix.

  1. Why PCA? My concern is that PCA is maximizing the variance and I am not sure this is exactly what we want.

  2. Are there other ways to choose the weights of this portfolio (apart from equally weighted or value weighted)? I am thinking, for instance, of minimizing the portfolio that is long a stock and short all others. Or I could use ANOVA.

  • 1
    $\begingroup$ Why do you care about beta in the first place? That will have a big effect on what to do. I suggest that there are reasons not to care about beta: portfolioprobe.com/2011/02/08/… $\endgroup$ Dec 30, 2011 at 10:23
  • $\begingroup$ @PatrickBurns Previous comments by you and the OP have been incorporated into the question above. $\endgroup$ Jan 1, 2012 at 0:11

2 Answers 2


Yes, the weights of the first eigenvector of a covariance matrix represent the market factor and also the largest source of systematic risk (variation of returns).

Why PCA? Well, PCA simply identifies the eigenvector that maximally explains the variance of the system. It turns out that this is the "market factor" - i.e. the tendency of securities to rise and fall together as an asset class. Why is this the market factor? If you examine the weights (factor loadings) of the first eigenvector in a histogram you will find they are generally all of the same sign whereas this is not the case for any of the subsequent eigenvectors (which represent sectors or style factors, that is to say other sources of systematic risk). In other words, it is empirically the case that there is a dominant systematic factor called the equity risk premium explaining the variance of returns.

Why is it empirically the case? Because macro variables such as monetary, fiscal policy, growth expectations, political risk, regulatory risk and other factors influence the returns of all stocks (although to varying degrees as measured by the security beta).

When you say "My concern is that PCA is maximizing the variance", it is more accurate to say that PCA is identifying that linear combination of security weights (i.e. an eigenvector) that has the highest variance which we call the market factor because the weights are generally the same sign. A portfolio constructed using these weights - an eigenportfolio - certainly would have high variance -- in fact it's the set of portfolio weights that produces the most volatility vs. any other set of portfolio weights since it loaded on the primary source of systematic risk. (As a side note, you could form a eigenportfolio that has minimum variance by identifying a principal component with a low eigenvalue.)

Indeed, if you plot the return of a portfolio based on the weights of the "first eigenportfolio" you will see the returns are highly correlated (but not equivalent to) the index itself as in the chart below from Marco Avellenada's paper:

enter image description here

There is a distinction between the market portfolio and the market factor described above. The "market portfolio" in the CAPM sense is the sum of all assets in the universe which by definition is market-cap weighted (since all assets are held in aggregate). Beta attempts to estimate the covariance of a security with respect to the market portfolio where the S&P or MSCI is a typical proxy. The market portfolio is fully diversified of idiosyncratic risk so its variance is explained by exposures to systematic risk only. The conceptual connection is that the Beta is the measure of the systematic risk or common source of variation in returns. Therefore, the first eigenvector (which identifies that linear combination of assets of maximal variance) is sometimes used as a proxy for the market portfolio. Also, a given security's "loadings" on the first principal component is one way of measuring the security's beta to the market factor.

On the second question, another way to choose the weights of the basket is market-cap weighted in line with the the theoretical definition of market portfolio. However, it's not clear what your goal is to "select weights of the basket". You correctly point out that you would simply be identifying a portfolio with pure exposure to systematic risk where idiosyncratic risk is diversified away.

  • $\begingroup$ I still think it is very accurate to say that "PCA is maximizing the variance". PCA weights can be found by formulating the problem as a max of the variance under constaint $\endgroup$ Feb 10, 2012 at 10:16
  • 1
    $\begingroup$ fyi - updated my answer to include the chart and some add'l perspective for posterity $\endgroup$ Apr 8, 2012 at 21:48
  • $\begingroup$ When you say "Why is this the market factor? If you examine the weights (factor loadings) of the first eigenvector in a histogram you will find they are generally all of the same sign whereas this is not the case for any of the subsequent eigenvectors", I am not convinced by this argument. I agree that the among all PCA components, the first one is the most representative of the market, but maybe there is another set of weights that is better than this one. $\endgroup$ Apr 18, 2012 at 4:52
  • $\begingroup$ Thank you for enhancing your answer, I appreciate reading it a lot. $\endgroup$ Apr 18, 2012 at 4:55

Look at the assets as copies of each other (say commonality $f$) with a bit of noise added to each asset. When you average out the assets you retain $f$ and noise goes to 0. $f$ is the "common part" embedded in each asset - this is what you need to know to pin down at once all the asset returns (to a reasonable extent) and thus this is what PCA abstracts out.

This lends itself to the natural interpretation of "market portfolio" - i.e. that which is embedded roughly in equal measure in all assets.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.