Advances in inflation forecasting with Random Forests

Why a new model?

In recent years, machine learning techniques have become increasingly popular across various scientific fields – including econometrics and statistics. New methods promise speedy processing of large amounts of data and the discovery of concealed relationships. For a long time, however, artificial intelligence algorithms were regarded as a "black box" whose calculations could neither be inspected nor edited. New techniques make the mechanics of the algorithms visible. On top of that, the calculation methods can also be used so flexibly that transparency of the forecast is enhanced.

For these reasons, we are introducing a purely data-driven model that uses two machine learning techniques: The Least Absolute Shrinkage and Selection Operator (LASSO) in combination with Random Forest models. Recent academic studies (such as this paper ) have shown that a combination of these methods have improved forecasts overall.

This model processes around 75 independent time series – and their numerous time lags – and estimates the main components of inflation with a forecast horizon of up to 12 months. Despite the high forecasting quality of these methods, the pure data model will continue to form only one, albeit important, component of our inflation forecasts. We supplement the data model with structural calculations and empirical values. The longer the forecasting horizon, the greater the importance of structural considerations.

For full text see attached PDF-Version.