Testing Forecasting Properties of Different Approaches to Interval Forecasting (Using the Example of Inflation in Russia)

Abstract

The paper contains the main results of the study of the quality of interval forecasts based on quantile regression and smooth quantile regression models relative to interval forecasts based on OLS models for Russian inflation. All models in the work were built from the logic of the possibility of forecasting in real-time, in connection with which the authors used inflation that was not cleared of seasonality in the expression month to the same month of the previous year. For each forecast horizon, the authors developed a separate model from lagged values of inflation and regressors so that the known values of lagged variables were always used on the right side of the equation.
The authors tested a wide range of specifications – from very compact autoregressive models with one or two lags to wide ones with three additional regressors in addition to inflation lags. The best quality from the point of view of CRPS (Continuous Ranked Probability Score) metrics typical for such a task were the specifications that included the only additional variable – the USD/RUB exchange rate. As the confidence interval narrowed and at longer horizons, quantile and smooth quantile models became increasingly better relative to OLS models according to the CRPS metric.
According to the authors, in the presence of inflation targeting, qualitative inflation forecasts can be used by the Bank of Russia in conducting monetary policy when forecasting future inflation dynamics and stress-testing the Russian economy. In general, this can help increase the economic agents’ confidence in the Bank of Russia within the concept of rational expectations