9.4.2 Empirical Evidence on DSGE Models

9.4.2 Empirical Evidence on DSGE Models

The key strength of DSGE models is their use of internally consistent model forecasts that facilitates inclusion of conditioning information such as current and future policy decisions. The jury is still out, however, on how helpful DSGE models are in terms of their ability to improve the accuracy of more ad hoc economic forecasts. In part this reflects the DSGE models’ disadvantages due to their use of representations of the economy that inevitably are highly simplified. If cross-equation restrictions imposed by DSGE models are rejected by the data, this will bias the forecasts. Countering such biases, DSGE models have the potential to reduce the effect of parameter estimation error.

In their empirical analysis, Del Negro and Schorfheide (2013) find that short-run forecasts of output growth and inflation from the Blue Chip survey are more precise than DSGE forecasts generated by a Smets and Wouters (2003) type model. At longer horizons between five and eight quarters, however, the DSGE forecasts of output growth are more precise than the corresponding survey forecasts. In turn Blue Chip interest rate forecasts are better than DSGE forecasts at horizons between one and eight quarters.

Ghent (2009) examines different DSGE models and their implied prior distributions in a forecasting exercise involving US output, investment, hours worked, and consumption. Despite considering a wide variety of models with different implications for the relationships between the variables, Ghent finds that the models yield Bayesian VARs with very similar forecast performance, although with improvements over an unconstrained VAR model as well as a simple BVAR with Minnesota priors.

Edge and Gürkaynak (2010) use Mincer–Zarnowitz regressions of outcomes on an intercept and the DSGE forecasts. If forecasts are unbiased, the intercept should equal 0 and the slope should equal 1; see chapter 15 for more details. For inflation forecasts they find that the intercept (α) is significantly positive, while the slope coefficient on the DSGE forecast $(\beta )$ is significantly smaller than 1, indicating that the forecasts are biased. A similar pattern emerges for output growth forecasts although here the evidence of bias is weaker in a statistical sense. For interest rate forecasts, the slope coefficient on the DSGE inflation forecasts exceeds 1 at one through six quarter forecast horizons and the intercept is negative and significantly different from 0 for the shortest horizons.

Summarizing the broader literature, Del Negro and Schorfheide (2013) conclude that “the empirical evidence in the literature suggests that DSGE model forecasts are comparable to standard autoregressive or vector autoregressive models but can be dominated by more sophisticated univariate or multivariate time-series models.”

Del Negro and Schorfheide (2013) argue that some of the inferior short-run predictive performance of the DSGE models may be due to their use of a narrow information set compared with survey forecasts which can make use of a larger set of variables and more updated information. They point at several ways for improving DSGE forecasts by better incorporating external information. To deal with the informational disadvantage that these models have vis-à-vis survey forecasts, Del Negro and Schorfheide suggest using observable long-run expectations data when estimating the model parameters. Such long-run inflation expectations can be used to determine the target inflation rate by modifying the central bank’s interest rate feedback rule and augmenting the model with a stationary time-varying inflation target.

An alternative strategy is to use external nowcasts from survey data to improve the model’s estimates of current and future states as well as to estimate model parameters. Nowcasts from surveys embed substantial amounts of information and so can be used to summarize broad sets of information. Specifically, at the one-period forecast horizon Del Negro and Schorfheide propose to modify $p(s_T,s_{T+1}|\theta ,y_1,...,y_T)$ in (9.45) by $p(s_T,s_{T+1}|\theta ,y_1,...,y_T,z_{T+1})$ , where $z_{T+1}$ is the one-period Blue Chip nowcasts which are available prior to the publication of ${\gamma }_{T+1}$

Forecasting with DSGE models remains a highly active and promising area of research with the potential for coming up with ways to improve how these models incorporate extraneous information while simultaneously incorporating constraints from economic theory in a way that can help reduce estimation error and guide the use of these models in providing feedback for policy advice.