Inflation forecasting plays a central role in monetary policy formulation. At the same time, recent international empirical evidence suggests that with the decline in inflation of recent years, the joint dynamics of this variable and its potential predictors has changed and inflation has become more unpredictable. Using a univariate model as a benchmark, we evaluate the predictive capacity of certain causal models linked to different inflation theories, such as the Phillips Curve and a monetary VAR. We also analyze the predictive power of models that use factors that combine the overall variability of a large number of business cycle time series as predictors. We compare their relative performance using a set of parametric and non-parametric tests proposed by Diebold and Mariano (1995). Although the univariate model performs best, as the forecast horizon lengthens, multivariate models performance improves. In particular, a monetary VAR performs better than the univariate ARMA model in the case of a one-year horizon. Nevertheless, when tests are calculated to evaluate the statistical significance of differences in the predictive capacity of models, taking a univariate ARMA model as a benchmark, differences are not statistically significant. Finally, estimated models are pooled to forecast inflation. Some of the forecast combinations outperform the best individual forecast over a one-year horizon. Taking into account that a one year-horizon is relevant for economic policy decisions, the possibility of combining both univariate and multivariate models for forecasting purpose is interesting, because it it can also be helpful to answer specific economic policy questions.
Argentina inflation forecast multivariate models pooling univariate models