Forecasting Albanian time series with linear and nonlinear univariate models

Abstract

Albanian economic time series show irregular patterns since the 1990s that may affect economic analyses with linear methods. The purpose of this study is to assess the ability of nonlinear methods in producing forecasts that could improve upon univariate linear models. The latter are represented by the classic autoregressive (AR) technique, which is regularly used as a benchmark in forecasting. The nonlinear family is represented by two methods, i) the logistic smooth transition autoregressive (LSTAR) model as a special form of the time-varying parameter method, and ii) the nonparametric artificial neural networks (ANN) that mimic the brain’s problem solving process. Our analysis focuses on four basic economic indicators – the CPI prices, GDP, the T-bill interest rate and the lek exchange rate – that are commonly used in various macroeconomic models. Comparing the forecast ability of the models in 1, 4 and 8 quarters ahead, we find that nonlinear methods rank on the top for more than 75 percent of the out-of-sample forecasts, led by the feed-forward artificial neural networks. Although the loss differential between linear and nonlinear model forecasts is often found not statistically significant by the Diebold-Mariano test, our results suggest that it can be worth trying various alternatives beyond the linear estimation framework.