Monetary aggregation and GDP forecasting in a support vector regression framework

The main objective of this study is to compare the forecasting performance of the Simple Sum and Divisia monetary aggregates in terms of the U.S. GDP, by employing an SVR approach. We use two alternative Divisia aggregates, the series produced by the Center for Financial Stability (CFS Divisia) and those produced by the Federal Reserve Bank of St. Louis (MSI Divisia). The use of the Divisia monetary aggregates in our study is justified by the Barnett critique that reveals the superiority of the Divisia against the simple sum monetary aggregates. This superiority is illustrated by the fact that the simple sum aggregates are not consistent with the microeconomic aggregation theory or the index number theory. Yet, the simple sum aggregates are widely used by central banks, economists and researchers even today. Five levels of monetary aggregation are used for each monetary aggregate, from the narrowest, to the widest, namely M1, M2M, M2, MZM and ALL. The empirical analysis is conducted within a machine learning framework employing a Support Vector Regression (SVR) model, as proposed by Vapnik in 1995, which uses an ε-insensitive loss function for solving regression problems, equipped with two kernels: the linear and the radial basis function (RBF) kernel. We also use for comparison reasons the widely employed VAR methodology. One and two quarters ahead forecasting windows are used. Our training data span the period from 1999Q4 to 2009Q3 and the out-of-sample forecasts cover the period 2009Q4 to 2012Q3. Our approach includes both static and dynamic forecasting models and based on Barnett’s critique we expect that our results will show that the Divisia monetary aggregates are superior to the Simple Sum monetary aggregates in terms of U.S. GDP forecasting.