The European Central Bank published Working Paper Series No 3052 proposing a Bayesian machine-learning structural inflation model that allows non-linear shock transmission at all impulse-response horizons. Applying the approach to euro area data, the authors find that inflation reacts disproportionately to large shocks while small shocks often generate no statistically significant response, with the strongest size-dependent effects linked to energy shocks, suggesting exceptionally large cost-push shocks may call for a differentiated monetary policy response. The model combines a Bayesian vector autoregression with Bayesian Additive Regression Trees and a non-linear structural factor model, identifying shocks via sign, zero and magnitude restrictions. Using monthly data from 1996:01 to 2024:11 across 10 variables, the paper identifies four shocks (energy, global supply chains, domestic supply and demand) and reports that non-linearities are most pronounced for energy shocks, rising smoothly with shock size rather than through discrete regime shifts. Along the pricing chain, the size-dependence is more visible upstream (energy commodity prices and energy producer prices) and attenuates downstream for consumer prices, with peak effects occurring immediately for the synthetic energy commodity index, around six months for energy producer prices and around one year for headline HICP. A sub-sample exercise indicates energy shocks have become more persistent and stronger in the period from 2020:01 onwards, with peak inflation effects later and dissipation closer to two years than in the pre-2020 sample. In a recursive density-forecasting comparison against a linear Bayesian VAR benchmark, the proposed model is broadly comparable at one-month horizons and improves forecast accuracy at longer horizons, particularly for higher-inflation outcomes.
European Central Bank 2025-05-06
European Central Bank working paper develops a non-parametric structural inflation model and finds euro area inflation reacts disproportionately to large energy shocks
The European Central Bank's Working Paper Series No 3052 introduces a Bayesian machine-learning structural inflation model capturing non-linear shock transmission. It shows euro area inflation reacts disproportionately to large shocks, especially energy shocks, suggesting differentiated monetary policy may be needed. The model improves forecast accuracy over a linear Bayesian VAR benchmark, particularly for longer horizons and higher-inflation scenarios.