Short-Term CPI Inflation Forecasting: Probing with model combinations

Forecasts of CPI inflation are critical in many public policy areas and private business planning. Many alternative approaches for selecting CPI forecasting models have been proposed. The standard practice to CPI forecasting is to pursue a winner-take-all perspective by which, for each dataset, a single believed to be the best model is selected from a set of competing approaches. However, model combination methods are becoming a common alternative to using a single time series method. We propose and apply a flexible Bayesian model averaging (BMA) approach of CPI inflation models to mitigate conceptual uncertainty and improve the short-term out-of-sample forecasting accuracy. The model space includes novel machine learning and deep learning algorithms and traditional univariate seasonal time series methods. The empirical results on the United States and Euro Area data reveal that BMA increases the predictive accuracy of CPI inflation forecasts in short-term exercises. The reduced out-of-sample forecast errors of BMA may be explained by their flexibility and capacity to select models that capture the diversity and complexity of inflation determinants and to estimate model weights that reflect the out-of-sample accuracy of a model.