In- and Out-of-Sample Performance of Nonlinear Models in International Price Differential Forecasting in a Commodity Country Framework

This paper presents an analysis of a group of small commodity-exporting countries' price differentials relative to the US dollar. Using unrestricted self-exciting threshold autoregressive models (SETAR), we evaluate the sixteen national Consumer Price Indexes (CPI) differentials relative to the US dollar CPI. Out-of-sample forecast accuracy is estimated through calculation of mean absolute errors measures based on the monthly rolling window and recursive forecasts, and this estimation is extended to three additional models, namely a logistic smooth transition regression (LSTAR), an additive nonlinear autoregressive model (AAR), and a simple neural network model (NNET). Our preliminary results confirm the presence of some form of nonlinearity in most of the analyzed countries. The parsimonious AR(1) model does not appear to perform any worse than any nonlinear model in the rolling sample exercise. However, in terms of a long- run equilibrium driven by purchasing power parity, its validity is undermined by the results of the recursive estimates and the outcome of the Diebold-Mariano type tests, which favor generally the Heckscher commodity points theory. As a policy advice to commodity-exporting countries, we find no apparent reason to suggest commodity export price pegging as a generalized foreign exchange policy.