Inflation forecasting is one of the central issues in micro and macroeconomics. Standard forecasting methods tend to follow a "winner-take-all" approach by which, for each time series, a single believed to be the best method is chosen from a pool of competing models. This paper investigates the predictive accuracy of a metalearning strategy called Arbitrated Dynamic Ensemble (ADE) in inflation forecasting using United States data. The findings show that: i) the SARIMA model exhibits the best average rank relative to ADE and competing state-of-the-art model combination and metalearning methods; ii) the ADE methodology presents a better average rank compared to widely used model combination approaches, including the original Arbitrating approach, Stacking, Simple averaging, Fixed Share, or weighted adaptive combination of experts; iii) the ADE approach benefits from combining the base-learners as opposed to selecting the best forecasting model or using all experts; iv) the method is sensitive to the aggregation (weighting) mechanism.