A forecasting performance comparison of dynamic factor models based on static and dynamic methods

Abstract

We present a comparison of the forecasting performances of three Dynamic Factor Models on a large monthly data panel of macroeconomic and financial time series for the UE economy. The first model relies on static principal-component and was introduced by Stock and Watson (2002a, b). The second is based on generalized principal components and it was introduced by Forni, Hallin, Reichlin (2000, 2005). The last model has been recently proposed by Zaffaroni (2015, 2016). The data panel is split into two parts: the calibration sample, from February 1986 to December 2000, is used to select the most performing specification for each class of models in a insample environment, and the proper sample, from January 2001 to November 2015, is used to compare the performances of the selected models in an out-of-sample environment. The metholodogical approach is analogous to Forni, Giovannelli, Lippi and Soccorsi (2016), but also the size of the rolling window is empirically estimated in the calibration process to achieve more robustness. We find that, on the proper sample, the last model is the most performing for the Inflation. However, mixed evidencies appear over the proper sample for the Industrial Production.

Calibrating Dynamic Factor Models with Genetic Algorithms

“…We first analyse a dataset of 176 EU macroeconomic and financial time series and then we conduct the same study on a dataset of 115 US macroeconomic and financial time series. In both studies, the employment of genetic algorithm in the calibration procedure produces very good results and more significant than those achieved in similar studies, such as [1,2].Abstract. In this work, we address the problem of calibrating dynamic factor models for macroeconomic forecasting.…”

“…Hence they are more robust to misspecification than frequency-domain methods. Instead, a systematic comparison of the forecasting performances of SW, FHLR and FHLZ can be found only in [1,2]. [2] conducted a forecasting exercise on a US macroeconomic dataset, taking an autoregressive process of order 4 as a benchmark.…”

“…Hence, Forni et al concluded that, due to its more dynamical structure, FHLZ tends to be the best performing method in "stationary periods", but it loses ground during regime changes. [1] conducted a forecasting exercise on an EU macroeconomic dataset. The global settings of his exercise are basically the same as in [2], but also the length of the rolling window is suboptimally selected during the calibration process.…”

“…Since each model is characterized by several parameters to estimate, an exhaustive exploration of the parameter space would be computationally infeasible. In order to give a partial solution to this issue, in [1] and in [2] the calibration procedure is carried out in a naif fashion, i.e. an initial configuration for each parameter is randomly selected and then, for each parameter at a time, a predetermined range of values is tested while keeping the other parameters fixed.…”

“…In this work, we also compare the macroeconomic forecasting performance of the three selected dynamic Author Proof factor models on two datasets. The former (an EU macroeconomic and financial dataset) is the same employed in [1]. Instead, the latter (an US macroeconomic and financial dataset) is the same employed in [2].…”