CMStatistics 2016: Start Registration
View Submission - CFE
Title: Forecasting global temperature with time-series methods Authors:  Marco Lippi - Universita di Roma La Sapienza (Italy)
Umberto Triacca - University of L Aquila (Italy)
Alessandro Giovannelli - University of Rome Tor Vergata (Italy) [presenting]
Antonello Pasini - National Research Council (Italy)
Alessandro Attanasio - University of L Aquila (Italy)
Abstract: The impact of climate change on territories, ecosystems and humans has been dramatic in the last fifty years and is likely to become heavier in the next decades, this making modeling and forecasting climate indicators, Global Temperature in particular, of the utmost importance. We propose the use of time-series methods, which are only weakly based on physical knowledge but provide an efficient use of the data available. We study and compare the forecasting performance of two models. The first is a standard Vector AutoRegression (VAR), in which Global Temperature is predicted using past values of (1) Global Temperature itself, (2) the greenhouse gases radiative forcing, GHG-RF, (3) the Southern Oscillation Index, SOI. The second is a Large-Dimensional Dynamic Factor Model (DFM). We use the information contained in 140 time series of local temperatures, corresponding to a grid with spacial resolution of $2.5 \times 2.5$ degrees. Our main findings are: (a) The cointegrated VAR, including GHG-RF, and SOI, performs better than the Factor Model at all horizons from 1 to 10 years.(b) However, augmenting the data in the VAR with factors (FAVAR) we obtain a competitive model. Moreover, averaging the forecasts of the FAVAR and the cointegrated VAR we obtain the best results.