The role of survey-based expectations in real-time forecasting of US inflation

This paper performs a real-time forecasting exercise for US inflation from 1992Q1 to 2022Q2. We reinvestigate the literature on autoregressive (AR) inflation gap models - the deviation of inflation from long-run inflation expectations. The findings corroborate that, while simple models remain hard to beat, the multivariate extensions to the AR gap models can improve forecasting performance at short horizons. The results show that (i) forecast combination improves forecast accuracy over simpler models, (ii) aggregating survey measures, using dynamic principal components, improves forecast accuracy, (iii) and the additional information obtained from the error correction process between inflation and long run inflation expectations can improve forecasting performance. In spite of our models providing more accurate one-step ahead forecasts on average, fluctuation tests reveal that over unstable time periods - mainly during the GFC and the Covid-19 pandemic - the AR(1) benchmark performed better.