Title: The evaluation of statistical learning models in macroeconomic time series
Authors: Simone Tonini - Sant Anna School of Advanced Studies - Pisa (Italy) [presenting]
Abstract: Thanks to the great amount of available data a part of the econometric literature focused on the application of statistical learning models in macroeconomics forecasting. Despite the interesting results obtained on forecasting accuracy, few works check whether these models are also able to get a good representation of the true data generating process. This kind of analysis is not trivial, specially when few variables determine the outcome in contexts with multicollinearity and serially correlated data. Moreover, understanding the role of each variable on the data generating process is crucial when the purpose is to understand and replicate the evolution and the dynamics of certain complex economic phenomena. To clarify this point, a large simulation study is presented in order to compare the performances in forecasting, variable selection and coefficient estimation accuracy of a large set of models under several data generating process and signal-to-noise ratio. Furthermore, the role of the serial correlation in the formation of spurious correlations is studied and a serial-decorrelation pre-step is proposed, in order to evaluate possible improvements in variable selection and forecasting accuracy using the i.i.d. component of the time series data.