A0516
Title: Spectral-based variable selection of high-dimensional data for prediction of the El Nino/Southern Oscillation cycle
Authors: Alessandro Giovannelli - University of L'Aquila (Italy) [presenting]
Tommaso Proietti - University of Roma Tor Vergata (Italy)
Abstract: The El Nino/Southern Oscillation (ENSO) phenomenon is a key driver of interannual climate variability. A novel procedure is introduced based on large dynamic factor models (DFMs) to improve the prediction of sea surface temperatures in the four El Nino regions. A significant body of literature on DFMs addresses the selection of variables for factor estimation, which directly impacts predictive accuracy. Existing methodologies for constructing targeted principal components often rely on static correlation for variable selection. The approach departs from these methodologies by proposing a new selection procedure that screens variables based on significant correlation within the frequency band most relevant to the variable of interest. This method, known as dynamic correlation, assesses co-movements between the target variable and covariates within the frequency band of interest, capturing dynamic correlations over time. This methodology is applied to a high-dimensional dataset containing sea surface temperatures from the Nino regions, using El Nino 3.4 as the target variable. To evaluate the effectiveness of the procedure, a real-time exercise is conducted comparing results obtained using the targeted dataset against those obtained using all available series. Preliminary results indicate significant potential to improve ENSO prediction accuracy.
