Bottom-up Inflation Forecasting Using Machine Learning Methods

Abstract

Machine learning methods are comparable or – often – superior to econometric ones. Machine learning models are more accurate as the amount of data grows, and time series for each item of the consumer price index (CPI) basket are available. Still, most studies neglect bycomponent data / shirk bottom-up approach. Few studies forecast the CPI by aggregating forecasts of price indices for individual goods and services (using the bottom-up approach). Based on existing research, one cannot be certain that the aggregated CPI inflation forecast is more accurate compared to forecasting headline CPI. On Russian data, we show that, depending on the forecast horizon, a by-component aggregated inflation forecast can be up to 1.5 times more accurate. Even with simple machine learning models, e.g. gradient boosting or regularised regression, the advantage is statistically significant on horizons of up to six months. We forecast each component of the consumer inflation at each horizon with a separate model, regardless of the other components or the other horizons.