A1284
Title: The specification of a fractionally integrated factor model
Authors: Dominik Ammon - University of Regensburg (Germany) [presenting]
Tobias Hartl - Maastricht University (Netherlands)
Rolf Tschernig - Universitaet Regensburg (Germany)
Abstract: The purpose is to investigate a possible negative effect of unnecessary differencing of nonstationary panel data. A factor model framework is utilized where factors may exhibit fractional integration. Nonstationary factors become more visible when the number of time periods $T$ tend to infinity due to an increasing variance. To mitigate the extreme signal-to-noise ratio, the signals of the fractional integrated factors are adjusted. The analysis demonstrates that the common component can be reliably estimated using principal component analysis as the number of variables $N$ and time periods $T$ tends to infinity. Specifically focusing on approximate dynamic factor models, the standard model selection criteria are established to remain effective for nonstationary and fractionally integrated data, assuming the idiosyncratic component is asymptotically stationary. Compared to the prior study, wherein the variance of I(1) factors tended towards infinity, the visibility of the factors is decreased by diminishing signals. Building upon the signals, a detrimental effect of differencing is found, as it reduces the influence of nonstationary factors, rendering them practically undetectable.
