Averaging Impulse Responses Using Prediction Pools

Abstract

Macroeconomists construct impulse responses using many competing time series models and different statistical paradigms (Bayesian or frequentist). We adapt optimal linear prediction pools to efficiently combine impulse response estimators for the effects of the same economic shock from this vast class of possible models. We thus alleviate the need to choose one specific model, obtaining weights that are typically positive for more than one model. Three Monte Carlo simulations and two monetary shock empirical applications illustrate how the weights leverage the strengths of each model by (i) trading off properties of each model depending on variable, horizon, and application and (ii) accounting for the full predictive distribution rather than being restricted to specific moments.

Impulse Response Functions for Self-Exciting Nonlinear Models

“…Several features of the MA approach proposed here are worth highlighting. Our weights do not vary across the impulse response horizon [as they do in Ho et al (2023) and Li et al…”

“…The second strand uses model averaging as a solution to model uncertainty problems when computing IRFs [e.g., Ho et al (2023) and Li et al (2021)]. Ho et al (2023) averages over a variety of structural models using prediction pool in the spirit of Geweke and Amisano (2011); Li et al (2021) constructs averages using the Stein combination estimator proposed in Hansen (2016). These papers differ from ours in at least two dimensions.…”

Bayesian Local Projections

“…An early application of BLP to the study of monetary policy shocks has appeared in Miranda-Agrippino and Ricco (forthcoming) together with the replication codes. Ho, Lubik and Matthes (2021) include BLP alongside other models in prediction pools designed for the estimation of robust impulse response functions. The BLP methodology is also distributed within the econometric package of Canova and Ferroni (2020).…”