12.07.2015 Views

"Frontmatter". In: Analysis of Financial Time Series

"Frontmatter". In: Analysis of Financial Time Series

"Frontmatter". In: Analysis of Financial Time Series

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

FACTOR ANALYSIS 345matrix P to transform the factor model so that the common factors have nice interpretations.Such a transformation is equivalent to rotating the common factors in them-dimensional space. <strong>In</strong> fact, there are infinite possible factor rotations available.Kaiser (1958) proposes a varimax criterion to select the rotation that works well inmany applications. Denote the rotated matrix <strong>of</strong> factor loadings by L ∗ =[l ∗ ij ] andthe ith communality by c 2 i .Define ˜l ∗ ij= l ∗ ij /c i to be the rotated coefficients scaledby the (positive) square root <strong>of</strong> communalities. The varimax procedure selects theorthogonal matrix P that maximizes the quantityV = 1 k⎡(m∑ ∑⎣ k ( ˜l ij ∗ )4 − 1 k∑ki=1i=1j=1˜l ∗2ij) ⎤ 2⎦ .This complicated expression has a simple interpretation. Maximizing V correspondsto spreading out the squares <strong>of</strong> the loadings on each factor as much as possible.Consequently, the procedure is to find groups <strong>of</strong> large and negligible coefficientsin any column <strong>of</strong> the rotated matrix <strong>of</strong> factor loadings. <strong>In</strong> a real application, factorrotation is used to aid the interpretations <strong>of</strong> common factors. It may be helpful insome applications, but not so in others.8.8.3 ApplicationsGiven the data {r t } <strong>of</strong> asset returns, the factor analysis enables us to search for commonfactors that explain the variabilities <strong>of</strong> the returns. Since factor analysis assumesno serial correlations in the data, one should check the validity <strong>of</strong> this assumptionbefore using factor analysis. The multivariate Portmanteau statistics can be usedfor this purpose. If serial correlations are found, one can build a VARMA modelto remove the dynamic dependence in the data and apply the factor analysis to theresidual series. For many returns series, the correlation matrix <strong>of</strong> the residuals <strong>of</strong> alinear model is <strong>of</strong>ten very close to the correlation matrix <strong>of</strong> the original data. <strong>In</strong> thiscase, the effect <strong>of</strong> dynamic dependence on factor analysis is negligible.Example 8.8. Consider again the monthly log stock returns <strong>of</strong> IBM, Hewlett-Parkard, <strong>In</strong>tel, Merrill Lynch, and Morgan Stanley Dean Witter used in Example 8.7.To check the assumption <strong>of</strong> no serial correlations, we compute the Portmanteaustatistics and obtain Q 5 (1) = 34.28, Q 5 (4) = 114.30, and Q 5 (8) = 216.78. Comparedwith chi-squared distributions with 25, 100, and 200 degrees <strong>of</strong> freedom, thep values <strong>of</strong> these test statistics are 0.102, 0.156, and 0.198, respectively. Therefore,the assumption <strong>of</strong> no serial correlations cannot be rejected even at the 10% level.Table 8.9 shows the results <strong>of</strong> factor analysis based on the correlation matrix usingboth the principal component and maximum likelihood methods. We assume that thenumber <strong>of</strong> common factors is 2, which is reasonable according to the principal componentanalysis <strong>of</strong> Example 8.7. From the table, the factor analysis reveals severalinteresting findings:

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!