Correct Comparison of Predictive Features of Machine Learning Models: The Case of Forecasting Inflation Rates in Siberia

This paper compares the quality of forecasts using machine learning methods and those produced by traditional models using the same information set via the example of forecasts of inflation rates in Siberian regions. We start with forecasts of regional inflation for various periods using several machine learning and benchmark methods. We then combine the forecasts produced by machine learning methods and weigh them against the resulting quality metrics. Finally, we compare the quality metrics with our benchmarks and confirm the robustness of the results using the Diebold–Mariano test. Based on the results of our study, we conclude that machine learning methods work better than benchmarks for most inflation time series longer than one year, in contrast to the forecasts for one–three quarters ahead. However, it is necessary to assess the quality of forecasting with machine learning methods for each region in advance to determine whether it makes sense to use them over traditional econometric tools. Forecasting with combined machine learning models appears to be preferable to any single model in most cases.