Academic journal article UTMS Journal of Economics

Bayesian Forecasts Combination to Improve the Romanian Inflation Predictions Based on Econometric Models

Academic journal article UTMS Journal of Economics

Bayesian Forecasts Combination to Improve the Romanian Inflation Predictions Based on Econometric Models

Article excerpt

INTRODUCTION

The inflation targeting needs accurate forecasts of the inflation rate in order to have a successful implementation of the monetary policy. Therefore, it is necessary to know some predictions methodologies that describe specific evolution of the inflation rate. Most of the central banks do not use only some individual models, but also suitable combined forecasts based on these models. In literature many researchers established that the combination of individual models forecasts outperform the predictions based on a single model. In the context of the economic crisis, Julio Roman and Bratu Simionescu (2013) showed that the reduction of forecasts uncertainty should be one of the major objective of experts in forecasting. The lower uncertainty of forecasts will improve the decisional process at macroeconomic level, but Terceno and Vigier (2011) showed that the business decisions are also improved.

An important review regarding the forecasts combination was made by Timmermann (2006). Diebold and Pauly (1990) have proposed a Bayesian shrinkage methodology in order to include prior information for improving the predictive accuracy of the combined forecasts. Authors like Wright (2008) or Koop and Potter (2003) used as prior mean zero-weights or equal-weights. Gomez, Gonzalez and Melo (2012) proposed a rolling window estimation method for co-integrated data series of order one in order to calculate the Bayesian weights.

In Romania, the National Bank uses a complex model for short and medium-mn predictions. However, the central bank did not make a combination based on Bayesian approach in order to improve the accuracy of its forecasts. Therefore, the object of this article is to make predictions of the inflation rate in Romania using the own econometric models, but also utilizing the Bayesian combination technique in order to improve the accuracy of individual expectations. After a brief description of the methodology, an empirical application is proposed for The Romanian inflation rate forecasts. All the individual models are valid econometric models of the inflation rate, being proposed by us. In this study we will use a prior mean that takes into account the forecasts based on Dobrescu macro-model, which is actually the first international model recognized for the Romanian economy.

METHODS

We consider a number of m h-step-ahead forecasts of the variable denoted by yt: ft/t-h'-'ft/t-h- Granger and Ramanathan (1984) proposed the following forecasts combination:

... (1)

CC = (ccq, CC],..., CCm) - vector of regression coefficients

ft/t-h = ( 1- fi/t-h' *** 'ftft-h) " vector that contains the intercept and a number of m forecasts (this vector dimension is m+1)

The intercept is introduced to ensure that the bias correction of the combined prediction is optimally determined.

Diebold and Pauly (1990) developed a method for introducing prior information in the regression of forecasts combination by using the g-prior model proposed by Zellner (1986). The model error is independently, normally and identically distributed of average 0 and variance σ2. Moreover, it is used a natural conjugate normal-gamma prior:

... (2)

The form of likelihood function is:

... (3)

The marginal posterior of a is:

... (4)

The marginal posterior mean is:

... (5)

where:

...

Diebold and Pauly (1990) showed the validity of the following relationship for g- prior analysis (M=gF'F):

... (6)

ge [0 ,00) is the shrinkage parameter. This parameter controls the relative weight between the maximum likelihood estimator and the prior mean in the posterior mean.

Wright (2008) utilized zero weight as the prior mean, while Diebold and Pauly (1990) recommended the equal weights. Geweke and Whiteman (2006) specified the prior distribution in Bayesian forecasting by including forecasters (experts) information. In this study we will use as prior weights the estimated parameters of the regression between the forecasters' h-step predictions and the forecasts based on different econometric models. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.