engleski

On the predictive power of statistical factor models

To obtain a risk decomposition of a large number of financial assets, latent factors are estimated from the data with an assumption that the estimated model will hold in the future. However, due to the inherent nonstationarity of financial systems, the estimated factor structures can only be assumed to hold over relatively short time frames, thus making traditional statistical factor models perform poorly outside those frames. In this paper we consider stastical methods for estimating latent factors from asset return time series. The considered framework includes two main steps: (i) estimating latent factors from observed asset returns (decomposition), and (ii) reconstruction of asset returns based on estimated factors. The unsupervised decomposition step is carried on a lookback window of recent past time series data and implies a factor model which is believed to hold in the future. The main goal of a model estimated on past data is to generalize well outside the estimation window (out-of-time) and explain the largest portion of return variance in a future period defined by the investment horizon. However, even when testing out-of-time performance using future return realizations, the realizations of latent factors are unknown. This problem can be resolved by using the estimated factor structure (precisely, the (i) decomposition step) to extract the implied factor realizations from the out-of- time returns. By reconstructing the out-of-time returns in this way, we measure the ability of the factor model to account for the observed returns. Given the proposed framework, we compare the performance of standard statistical approaches based on principal components with a clustering-based approach which recovers sparse factor structures, as given in Fig. 1 (supplement). In addition, we build optimized factor-based portfolios using estimated factor structures. Our results, obtained using a well diversified asset universe, suggest that methods which may yield optimal results in-sample tend to generalize poorly outside the estimation window, and that the proposed clustering-based method demonstrates better performance in this regard.

Factor models ; Forecasting ; Finance

nije evidentirano