A1442
Title: 3D-PCA in foreign exchange markets
Authors: Moritz Dauber - University of Innsbruck (Austria) [presenting]
Abstract: The method of 3D-PCA is applied to a cross-section of currency portfolios to analyze the driving forces of risk premia in currency markets. 3D-PCA is a dimension reduction technique to extract (latent) factors based on a tensor model decomposition. One strong advantage of the 3D-PCA estimation over standard PCA is its robustness, even in short samples, leading to a particularly superior out-of-sample performance of the derived factors. A cross-section of currency portfolios is formed based on univariate sorts of various characteristics, such as the forward discount, the real exchange rate, momentum or external imbalances, and it is found that the latent factors extracted from 3D-PCA outperform standard PCA factors in terms of cross-sectional pricing errors as well as Sharpe ratios, particularly out-of-sample. Besides that, 3D-PCA allows conclusions regarding the economic interpretation of the (latent) factors to be drawn. Namely, the building blocks of the factors exhibit typical level, slope and curvature patterns of the underlying characteristics. A comparison to well-established factors from the literature shows that 3D-PCA factors perform reasonably well and yield overall similar pricing errors and Sharpe ratios.
