FORECASTING INFLATION WITH ANN MODELS

VICENTE RIOS IBÁÑEZ*

Msc. Macroeconomic Policy and Financial Markets (BGSE)

 BANCO CENTRAL DEL PARAGUAY

DOCUMENTOS DE TRABAJO

CENTRAL BANK OF PARAGUAY

WORKING PAPERS

December 2010

 *Se agradecen los consejos, opiniones y ayuda prestada por los compañeros Gustavo Biedermann, Victor Ruiz, Dario Rojas, Jazmín Gustale y Carlino Velázquez. Las opiniones y errores que puedan surgir en este documento, son de exclusiva responsabilidad del autor y no compromete la posición institucional del Banco Central del Paraguay.

ABSTRACT

In this research I investigate the alternative methodology of training artificial neural networks models with the early stopping procedure and I analyze their outcomes in terms of accuracy when forecasting monthly Paraguayan inflation time series. The results show that despite of neural network modelling being a competitive alternative to classical linear modelling it doesn‟t improve the overall forecast performance of best ARMA specifications selected through common in-sample estimation procedures, in a set of four control subsamples of 24 months each, ranging from 2002:04 to 2010:04. However, it is also a remarkable feature of all the checks performed in this research, that artificial neural network models outperform ARMA specifications in 24-steps-ahead horizon forecasts in all the subsamples of control.

INTRODUCTION

ANN modelling has gained attention in past years as an attractive technique for estimation and forecasting in economics. The main advantage of the artificial neural networks models is that they are free from the assumption of linearity that is commonly imposed in order to make the traditional methods tractable. Moreover, as Hornik and M-Stinchcombe and H.White (1989) Franses and Van Dijk (2000) show, neural networks are universal approximators and they can fit arbitrarily well any complex function by increasing the number of layers and neurons of the network.

Recent literature in ANN models show the great capability of ANN models in identifying behavior patterns, especially, non-linear ones, which allows the detection of non-linear dynamics and perform high accuracy forecasts. Nakamura (2005) shows that ANN methods outperform results obtained from linear autoregressive models. Moreover, Paul McNelis and Peter McAdamn (2004) show that non-linear Phillip curve specifications based on thick NN models can be competitive with the linear specification and that they perform better in periods of structural change. NN modeling also has the capability, as it has been shown in Tckaz (1999), to recognize and to model non-typical observations such as outlier behaviors or changes in the level, showing notorious advantages from what linear models can do in uncertain environments.

This document investigates alternative methods of forecasting and evaluates their performance in order to improve and complement inflation forecast exercises of Paraguay‟s Central Bank. The alternative approach purposed here constitutes a powerful alternative in regression standard techniques to model and forecast time series. In this research, I apply neural network-based trained models to: i) perform connections between the past and present values of the inflation time series and ii) to extract the structures and relations driving the information system associated to inflation.

In this research I proceed in the following way:

I first estimate different Autoregressive Moving Average (ARMA) models for different subsamples and I determine the most competitive specifications according to different information and goodness of fit criteria such as the Akaike‟s Information Criteria(AIC), the Bayesian Information Criteria(BIC), maximum likelihood (ML) and the minimum Root Mean Squared Error (RMSE).

Second, for each subsample I built and train a neural network to learn about each partition of the data with the early stopping procedure documented by Nakamura (2005) that will compete with the best ARMA(p,q) model specifications. I run a double loop in the MATLAB programming environment for lags and layers so that the training takes place accounting for all possible network configurations minimizing the in-sample RMSE for different configurations of lags and layers.

Third, I compute the out of sample root mean squared forecast error (RMSFE) for each subsample of interest for both, the ARMA(p,q) models specifications and the ANNs one. Finally, I analyze the predictive accuracy in each of them and conclude whether ANN models are competitive with ARMA(p,q) models or not.

My main finding is that Artificial Neural Networks (ANN) trained to forecast monthly inflation rate series in Paraguay with the early stopping procedure and with the learning process based on Levenger-Marquardt algorithm are a) competitive with ARMA(p,q) models but b) do not outperform univariate ARMA(P,Q) models for the all set of check subsamples analyzed according to the RMSFE measure. The later, for the case of Paraguay‟s inflationary rate process contradicts the intuition of superiority provided by in Tckaz (1999), Paul McNelis and Peter McAdamn (2004) Nakamura (2005), Haider and Hanif (2009), Manfred Esquivel (2009) among others, at least, when using macroeconomic Paraguayan data.

Forecasting Inflation With ANN Models - BCP - PortalGuarani by portalguarani

Fuente: BANCO CENTRAL DEL PARAGUAY