VAR 모형과 Bayesian VAR 모형의 물동량 예측력 비교 연구

Abstract

This paper presents a comparison of forecasting trade volume between general and Bayesian VAR(vector autoregressive) approaches. There exist a variety of investigation to find a unique relationship between trade volume and economic indicators on the basis of economic theory or properties of the dataset. In the contents of many approaches to the trade volume, interest focused on finding effective factors through the various evaluation methods such as long-run relationship or factor analysis. However, among these are the testing procedure, the presence of regression based causality and cointegration only. Alternatively, we propose multivariate approaches, general VAR and Bayesian VAR, to forecast trade volume including macroeconomic determinants such as foreign exchange rate and Baltic Dry-bulk Index(hereafter BDI) and show forecasting power between the frameworks. Specifically we perform in-sample and out-of-sample forecasts using general VAR and Bayesian VAR in this investigation. As a first step, we test stationarity of the series using parametric and nonparametric methods. All the series span the period January 2000 to April 2012 and show nonstationary of ADF(augmented Dickey-Fuller), KPSS(Kwiatkowski, Phillips, Schmidt and Shin) and LM(Leybourne and McCabe) tests in the level. When the series are first-differenced, the results shows stationary. These are also called unit root processes in the level, and processes integrated of order 1. We then perform the Johansen cointegration test among the variables but it shows insignificant, which is consistent with previous results that include global financial crisis periods in 2008. Macroeconomic practitioners frequently work with multivariate nonstationary time series. However, the majority face how to consider the stationarity issues as well as model specifications. Regarding the stationarity, Bayesian methods have become increasingly popular as a way of overcoming stationarity problem of the series. That is, Bayesian VAR can circumvent the information loss because the approach is able to estimate the model without the restrictions. We perform in-sample and out-of-sample forecasts to compare the power between general and Bayesian approaches. For comparing the performance between general and Bayesina VAR forecasts we first estimate mean-based error to the variables and then estimate such as MAFE (Mean Absolute Forecast Error), RMSFE (Root Mean Squared Forecast Error), and RMSPFE (Root Mean Squared Percentage Forecast Error) to evaluate the forecasting power of the models. Forecasting power is measured using the ration of Bayesian VAR against general approach and, while there is a little similarity with in-sample forecasts, its economic value is significantly different. In particular, the overall estimated results from error-based analysis to the out-of sample forecast imply that Bayesian VAR is superior to the general VAR forecasts in trade volume.