Title: Factor analysis for data with heterogenous blocks
Authors: Tatyana Krivobokova - University of Vienna (Austria) [presenting]
Abstract: It has been often empirically observed that including more variables into factor-based forecasting models may worsen the prediction considerably. We examine this issue assuming a factor model, which consists of heterogeneous and possibly dependent blocks of variables. We identify settings that cause the poor forecasting performance of such factor models estimated by principle component analysis (PCA) and suggest a simple modification of the standard PCA, called blocked PCA (bPCA), that leads to the proper identification of factors and prediction. A simulation study and real data analysis illustrate our findings.