Do German economic research institutes publish efficient growth and inflation forecasts? A Bayesian analysis

ABSTRACT

We use Bayesian additive regression trees to reexamine the efficiency of growth and inflation forecasts for Germany. To this end, we use forecasts of four leading German economic research institutes for the sample period from 1970 to 2016. We reject the strong form of forecast efficiency and find evidence against the weak form of forecast efficiency for longer-term growth and longer-term inflation forecasts. We cannot reject weak efficiency of short-term growth and inflation forecasts and of forecasts disaggregated at the institute level. We find that Bayesian additive regression trees perform significantly better than a standard linear efficiency-regression model in terms of forecast accuracy.

KEYWORDS:

• Forecast efficiency
• Bayesian modeling
• regression trees

JEL CLASSIFICATIONS:

• C53
• E31
• E32
• E37

Acknowledgments

We thank an anonymous associate editor and two reviewers for helpful comments. The usual disclaimer applies.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 For an early analysis of macroeconomic forecasting of European and German research institutes, see Ray [40].

2 Forecasts are often classified as weakly efficient if forecast errors have a mean of zero (unbiasedness) and are not autocorrelated. See, for example, Öller and Barot [33] and Timmermann [43]. On weak and strong forecast efficiency, see Nordhaus [32]. For a definition of weak and strong forecast efficiency and the concept of forecast rationality, see also Stekler [42].

3 For a detailed description of the BART model and an extension to Bayesian model averaging, see also the recent research by Hernández et al. [23].

4 The choice of a larger than the default value for the parameter for k is mainly motivated by model performance. We study the influence of the number and the structure of trees along with other robustness checks in Section 4.

5 We use the predicted CPI when available, however, in case of some forecasts the research institutes did not make a distinction between the CPI and the deflator of private consumption. In these cases, we assume that the institutes intended this absent distinction and use the available inflation data (for a similar approach, see [10]).

6 See also Behrens et al. [3]. The research institutes are (in alphabetical order): Deutsches Institut für Wirtschaftsforschung, Hamburger Weltwirtschaftsarchiv/-institut, Ifo Institut, and Institut für Weltwirtschaft.

7 Data source: https://www.destatis.de/EN/Publications/WirtschaftStatistik/WirtschaftStatistik.html

8 The German statistical office switched from GNP as the lead indicator for economic growth to GDP in 1992. This switch does not affect our data because we consider GDP forecasts and realizations throughout the whole sample period.

9 Results (not reported) of estimating Mincer-Zarnowitz regressions of forecast errors on a constant confirmed that the forecasts are unbiased on average.

10 Before 1999, we consider the exchange rate of the US dollar vis-à-vis the Deutsche Mark.

11 Coibion and Gorodnichenko [8] analyze in detail how serial correlation of the forecast error, along with the response of forecast errors to shocks and forecaster disagreement, can be used to test predictions of popular models of informational rigidities like sticky-information and noisy-information models.

12 Graphs showing partial dependence plots for all four types of forecasts and all predictors are available from the authors upon request.

13 The horizontal axis reflects the marginal distribution of the predictor series and does not reflect actual quantiles. For example, the 20%, 40%, 60%, and 80% quantiles of the annualized inflation rate are 1.05%, 1.60%, 1.89%, and 2.86%. Comparing these values with the bottom left graph illustrates that approximately 40% of the observed values lead to a significant (negative) response of $GD{P}_{q2}$.

14 In contrast, choosing a smaller value for the parameter ν would result in a highly ‘aggressive’ setup, which increases the possibility of overfitting (see [7] for a detailed discussion of the choice of hyperparameters).

15 For classification variants of random forests, see Behrens et al. [4].

16 The convergence statistics for the other forecast errors look similar and are not presented to save journal space. In order to mitigate the influence of autocorrelation during the iteration process, Kapelner and Bleich [25] propose to parallelize the sampling process after the burn-in period (vertical lines in the lower subplots).

17 We consider both, the lasso and ridge penalty and use ten-fold cross-validation for hyperparameter optimization.

18 An exception is $GD{P}_{q4}$, where the elastic net and random forests both lead to identical results.

19 The data were downloaded in July 2019 from the following internet page: https://www.policyuncertainty.com.

20 The data were downloaded in July 2019 from the following internet page: https://sites.google.com/view/jingcynthiawu/shadow-rates. The U.S. data start in 2009/1 and the German data start in 2004/9.

21 In this regard, it should be remembered that both models give similar results in terms of the significance of permutation tests (see Table 7).

Additional information

Funding

This work was supported by the German Science Foundation (Deutsche Forschungsgemeinschaft) (Project: Exploring the experience-expectation nexus in macroeconomic forecasting using computational text analysis and machine learning; Project number: 275693836).