Iterated Dynamic Model Averaging and application to inflation forecasting

Abstract

This manuscript presents a new forecasting methodology that builds upon the established Dynamic Model Averaging (DMA) method, termed the Iterated Dynamic Model Averaging (IDMA) algorithm. The IDMA algorithm works on the DMA framework by modifying its input parameters to optimize estimation on the training dataset, effectively selecting candidate predictor variables and calibrating key model parameters. To validate the forecasting efficacy of IDMA, we have conducted empirical analyses of IDMA and other benchmark models on inflation rate predictions. First, we present the forecast on the United States (US) as our primary result, followed by sensitivity analyses on various initial predictors and parameters. Subsequently, we expand the discussion to include other countries for further illustration. Finally, we reinforce our conclusions by conducting forecasts on simulated data through numerous replications. Our findings demonstrate that IDMA outperforms other benchmark models at yearly time horizon across diverse economic contexts and exhibits substantial robustness across varied initial configurations of predictors and parameters.