Files
Abstract
When applied to time series processes containing occasional level shifts, the logperiodogram (GPH) estimator often erroneously finds long memory. For a stationary short-memory process with a slowly varying level, I show that the GPH estimator is substantially biased, and I derive an approximation to this bias. The asymptotic bias lies on the (0,1) interval, and its exact value depends on the ratio of the expected number of level shifts to a user-defined bandwidth parameter. Using this result, I formulate the Modified GPH estimator, which has a markedly lower bias. I illustrate this new estimator via applications to soybean prices and stock market volatility.