On the macroeconomic forecasting performance of selected dynamic factor models

Abstract

In this doctoral thesis, we compare the forecasting performance of three dynamic factor models on macroeconomic and financial datasets. The purpose of the first two chapters is to provide an incremental contribution with respect to the body of literature comparing static versus dynamic factor models. Previous literature compares the forecasting performance of the static factor model SW (see Stock and M. W. Watson, 2002a, Stock and M. W. Watson, 2002b) against those of the dynamic factor model FHLR (see Forni, Hallin, Lippi, and Reichlin, 2000, Forni, Hallin, Lippi, and Reichlin, 2005). This work adds a third dynamic factor model, which is the recently published FHLZ (see Forni, Hallin, Lippi, and Zaffaroni, 2015, Forni, Hallin, Lippi, and Zaffaroni, 2016). In the third chapter, we compare the forecasting performance of two static factor models cast in a state-space form. In the first one, the conditional moments of the factors are estimated under proper hypothesis of linearity and gaussianity of the data. In the second one, the assumptions of linearity and gaussianity are relaxed for the estimation of the conditional moments of the factors. Chapter 1 presents an application of the three factor models (SW, FHLR and FHLZ) for forecasting purposes. It compares the pseudo real-time forecast performances of the three factor models against a benchmark AR(4) (an autoregressive process of order 4) over a dataset of 176 EU macroeconomic and financial time series. In this exercise, FHLZ generally outperforms all methods on the forecasting of the Consumer Price Index (CPI). Instead, no method seems to outperform the others in forecasting the Industrial Production (IP), but all dynamic factor models outperform the benchmark AR(4). Chapter 2 presents two applications on the same topic of the previous chapter. The most innovative part of these applications is that a genetic algorithm is employed to calibrate the three dynamic factor models. The first application exposed in this chapter employs the same dataset of Chapter 1. Instead, in the second application a dataset of 115 US macroeconomic and financial time series is employed. In this chapter, we show that FHLR tends globally to outperform the other methods on the real variables and that FHLZ tends globally to outperform the other methods on the nominal variables. As to EU dataset, in chapter 1 we found similar results for the CPI, but mixed evidences appeared for the IP. As to the US dataset, Forni, Giovannelli, et al., 2016 found similar but less significant results. Chapter 3 extends a previous study from Banbura and Modugno, 2014, by comparing the forecasting performance of a dynamic factor model cast in state-space form in which the conditional moments relative to the factors are estimated by means of the two following techniques: (i) Kalman filter: as in Banbura and Modugno, 2014, the conditional moments relative to the factors are estimated under the hypothesis that the data generating process (DGP) is linear and that the error terms follow a Gaussian distribution; (ii) Paticle Filter: in this case, the conditional moments relative to the factors are estimated in a more general framework, in which the DGP may be affected by sources of nonlinearity and in which the error terms may not follow a Gaussian distribution. Up to our knowledge, the estimation of the conditional moments of the factors by means of the Particle Filter has not been carried out yet. In this application, we employ the same Small dataset of 14 EU/US macroeconomic and financial time series from Banbura and Modugno, 2014. We show that the assumptions of linearity of the DGP and of a gaussian distribution for the error terms seems to hold in this macroeconomic setting. Hence, the estimation of the conditional moments of the factors by means of the Kalman Filter seems to be the more appropriate choice in macroeconomic forecasting. However, it is also possible that the particle filter may outperform in financial forecasting. As can be seen in Habibnia, 2017, it appears that accounting for the sources of nonlinearity in the DGP plays a more relevant role on forecasting financial variables.