Comparing Factor Models by Correlation Structure and Minimum-Variance Weights
Summary
The document asks how to compare sample, single-index, industry-factor, and principal-component factor models using their implied correlation matrices and minimum-variance portfolio weights. It describes plots from an analysis and proposes that a model whose correlation matrix resembles the sample matrix may explain the observed return relationships more closely. The author tentatively judges the industry-factor model to look most similar, but supplies no quantitative comparison or evidence beyond visual inspection.
The question also asks whether portfolio weights should resemble those computed from the sample covariance structure or whether another model is preferable. No answer or evaluation procedure is included, so the document does not establish a preferred model. A useful assessment would need criteria tied to the intended use, such as out-of-sample risk estimation or portfolio performance, along with attention to estimation uncertainty and weight stability. Visual similarity alone does not show that a model produces better minimum-variance portfolios.
Key ideas
- The document compares sample, single-index, industry-factor, and PCA models through correlation plots and minimum-variance weights.
- Visual similarity between a model correlation matrix and the sample matrix is proposed as a possible fit criterion.
- The industry-factor model is only tentatively judged closest by visual inspection.
- The document gives no evidence establishing which model yields superior minimum-variance portfolios.
- Model comparisons should be judged against an explicit objective and evaluated beyond in-sample visual resemblance.
Tags
Full text
# Interpreting different factor models w.r.t. correlation matrix and min variance portfolio weights # Interpreting different factor models w.r.t. correlation matrix and min variance portfolio weights ## Background In Eric Zivot's analysis of factor models he uses three models - The sample (.sample) - Single index model (.si) - Barra factor industry model (.ind) - PCA model (.pca) You can download his theory and notation macro models and theory and notation fundamental models as well as the analysis slides. I have a question on an analysis. It's from a second year PhD course and I'm a BSc student but I have spent this week trying to understand the theory and code and I think I'm on my way. ## Question I want to interpret two things: - The correlation matrix given by code `plotcorr()` i.e. in his slides page 7, 17, 36 for sample, .ind, .pca respectively. - The portfolio weights he has calculated for min variance portfolio, given by code `barplot(t(w.gmin.modelnamegoeshere))` i.e. page 7, 19, 39 for modelnames .sample, .ind, .pca respectively. So I wonder: Given these plots 4 correlation plots and 4 portfolio weights plot, are any model preferable to another? (Note that Zivot either forget or deliberately omit the .si models for plotcorr() and portfolio weights - it's in his code but didn't make it to the powerpoint.) ## My guess - For the first question, my guess is that you want to be close to the sample. Because we do cor(returns) for ".sample" ".si model" ".ind model" and ".pca model" and we wish to explain the return data with a model, so being similar to the sample is what we want. Here my eyes tell me .ind looks most like .sample and hence .ind model is better. - For the second question, my first guess was that you want to be close to the sample for the same reason as above. But then I hesitaded. What we do here to find - given corr(returns) and returns - a minimum variance portfolio. So the sample is a benchmark and we wish to have better risk adjusted return. Which is better of Zivot's models? I don't know.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.