A0868
Title: Bias-reducing penalization for the Whittle likelihood
Authors: Francesca Papagni - Free University of Bolzano (Italy) [presenting]
Davide Ferrari - University of Bolzano (Italy)
Greta Goracci - University of Bologna (Italy)
Abstract: Whittle likelihood estimation is a widely used and computationally efficient approach to approximate the Gaussian likelihood in the frequency domain. However, it produces biased parameter estimates when the samples are relatively small. The empirical adjustment is considered for the Whittle likelihood function, which is shown to reduce the size of the asymptotic bias of the resulting estimates. The validity of the approach is shown through asymptotic calculations and Monte Carlo experiments. The examples focus on long-memory models, for which the Whittle likelihood represents one of the most commonly used estimation methods. The numerical findings show that significant bias reductions in small samples characterize the adjusted Whittle estimates. As an illustration, the new methodology is applied to the analysis of the Southern Oscillation Index data based on a relatively short series, which is relevant for predicting year-to-year climate variation in the global climate leading to floods, droughts and other natural disasters.