Factor models are commonly used in financial applications to analyze portfolio risk and to decompose it to loadings of risk factors. A linear factor model often depends on a small number of carefully-chosen factors and it has been assumed that an arbitrary selection of factors does not yield a feasible factor model. We develop a statistical factor model, the random factor model, in which factors are chosen at random based on the random projection method. Random selection of factors has the important consequence that the factors are almost orthogonal with respect to each other. The developed random factor model is expected to preserve covariance between time-series. We derive probabilistic bounds for the accuracy of the random factor representation of time-series, their cross-correlations and covariances. As an application of the random factor model, we analyze reproduction of correlation coefficients in the well-diversified Russell 3,000 equity index using the random factor model. Comparison with the principal component analysis (PCA) shows that the random factor model requires significantly fewer factors to provide an equally accurate reproduction of correlation coefficients. This occurs despite the finding that PCA reproduces single equity return time-series more faithfully than the random factor model. Accuracy of a random factor model is not very sensitive to which particular set of randomly-chosen factors is used. A more general kind of universality of random factor models is also present: it does not much matter which particular method is used to construct the random factor model, accuracy of the resulting factor model is almost identical.
↧