European Central Bank working paper introduces historical-event decomposition to interpret local projection impulse responses

The European Central Bank published a working paper that proposes a new way to “open the black box” of local projections by decomposing impulse response estimates into a sum of contributions from specific historical observations. The approach is designed to show whether an estimated policy effect is broadly supported by the data or dominated by a small set of episodes, and to support narrative checking of what events are actually driving the estimate. For least squares local projections, the paper shows the underlying weights can be interpreted both as purified and standardised shocks and as proximity scores that measure similarity between the intervention being studied and past interventions in the sample. It extends the proximity-based weighting logic to many machine learning methods that remain linear combinations of outcomes, and proposes concentration metrics based on the share of total weights or contributions accounted for by the top 10% of observations. Empirical applications to monetary and fiscal shocks, global temperature shocks and the excess bond premium highlight that estimates can be heavily concentrated in identifiable episodes, including 1970s stagflation and Nixon-era interference with the Federal Reserve, World War II and the Korean War, the Mount Agung volcanic eruption and the Great Financial Crisis, with Random Forest-based estimates often producing sparser weight patterns intended to improve interpretability.