Predicting US inflation: Evidence from a new approach

In this paper, we further subject to empirical scrutiny the conclusion of Stock and Watson (1999) that commodity prices do not improve the traditional Phillips curve-based inflation forecasts. Thus, a multi-predictor framework for US inflation is constructed by augmenting the traditional Phillips curve with symmetric and asymmetric oil price changes. We show that the underlying predictors of US inflation exhibit persistence, endogeneity and conditional heteroscedasticity effects which have implications on forecast performance. Thus, we employ the Westerlund and Narayan (WN hereafter) (2012, 2015) estimator which allows for these effects in the predictive model. Also, we follow the linear multi-predictor set-up by Makin et al. (2014) which is an extension of the bivariate predictive model of WN (2012, 2015). Thereafter, we extend the former in order to construct a non-linear multi-predictor model that allows for asymmetries based on Shin et al. (2014) approach. Using historical monthly and quarterly data for relevant variables ranging from 1957 to 2017, we demonstrate that the oil price-based augmented Phillips curve will outperform the traditional version if the inherent effects in the predictors are captured in the predictive model. In addition, we also construct a Dynamic Model Averaging version for the augmented Phillips curve, as well as linear time-series models as Autoregressive Integrated Moving Average (ARIMA) and Fractionally Integrated versions (ARFIMA). The WN-based approach is found to outperform the alternative models that ignore the inherent effects. Our results are robust to different measures of inflation, data frequencies and multiple in-sample periods and forecast horizons.