On the power of underdifferencing and overdifferencing tests against nearly nonstationary alternatives

Abstract

The choice of the appropriate degree of integration is a very important question in ARIMA time series modeling. This choice is particularly difficult in the presence of either a nearly nonstationary autoregression or a fractionally integrated process. Via a Monte Carlo study we assess the size and power of MA, AR and spectral estimation tests in the presence of fractionally integrated, nearly nonstationary, and nearly noninvertible processes.

Keywords:

• ARIMA
• ARFIMA
• AR and MA unit roots