Loss function‐based evaluation of DSGE models

TLDR

The paper proposes a Bayesian econometric procedure to evaluate and compare DSGE models. The method introduces a reference model to handle misspecification, applies three loss functions to compare DSGE predictions with posterior population characteristics, and is used to compare a cash‑in‑advance and a portfolio‑adjustment‑cost model. The cash‑in‑advance model achieves higher posterior probability and better in‑sample fit, yet both models overpredict the negative output‑inflation correlation; only the portfolio‑adjustment‑cost model generates a positive real effect of money‑growth shocks and its impulse responses more closely match posterior means. © 2000 John Wiley & Sons, Ltd.

Abstract

Abstract In this paper we propose a Bayesian econometric procedure for the evaluation and comparison of DSGE models. Unlike in many previous econometric approaches we explicitly take into account the possibility that the DSGE models are misspecified and introduce a reference model to complete the model space. Three loss functions are proposed to assess the discrepancy between DSGE model predictions and an overall posterior distribution of population characteristics that the researcher is trying to match. The evaluation procedure is applied to the comparison of a standard cash‐in‐advance (CIA) and a portfolio adjustment cost (PAC) model. We find that the CIA model has higher posterior probability than the PAC model and achieves a better in‐sample time series fit. Both models overpredict the magnitude of the negative correlation between output growth and inflation. However, unlike the PAC model, the CIA model is not able to generate a positive real effect of money growth shocks on aggregate output. Overall, the impulse response dynamics of the PAC model resemble the posterior mean impulse response functions more closely than the responses of the CIA model. Copyright © 2000 John Wiley & Sons, Ltd.