Business-cycle reports and the efficiency of macroeconomic forecasts for Germany

ABSTRACT

We study the efficiency of growth and inflation forecasts published by three leading German economic research institutes during a period of time ranging from 1970 to 2017. To this end, we examine whether the information used by the research institutes when they formed their forecasts helps to explain the ex-post realized forecast errors. We identify the information that the research institutes used to set up their quantitative forecasts by applying computational-linguistics techniques to decompose the business-cycle reports published by the research institutes into various topics. Our results show that several topics have predictive value for the forecast errors.

KEYWORDS:

• Growth forecasts
• inflation forecasts
• forecast efficiency
• business-cycle reports

JEL CLASSIFICATION:

• C53
• E32
• E37

Acknowledgments

We thank an anonymous reviewer for helpful comments. The usual disclaimer applies.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Some researchers use the terms weak and strong forecast rationality rather than weak and strong forecast forecast-efficiency. See Stekler (2004).

² Another problem is that, when a researcher uses macroeconomic variables to represent the right-hand side variable, $x_t$, it is important to account for ex-post data revisions and possibly time-varying publication lags.

³ The research institutes are (in alphabetical order): Deutsches Institut für Wirtschaftsforschung, Ifo Institut, and Institut für Weltwirtschaft.

⁴ The basic idea is to map each word of the business-cycle reports via the skip-gram method in vector space and allows for a mathematical representation of the word sense. An LDA algorithm then extracts topics out of the obtained co-occurrence matrix. For an extensive explanation of this approach, see Panigrahi, Simhadri, and Bhattacharyya (2019) and Foltas (2020).

⁵ We train the topic model with additional business-cycle reports from the Hamburg Institute for the World Economy (HWWI) and the Joint Economic Diagnosis (JED, a group of research institutes) to increase the overall size of the text corpus and, thus, to make our results more robust. Because only relatively few forecasts are available for the HWWI and the JED, we do not use the results for the HWWI and the JED in our further analysis.

⁶ Our model extracts 24 topics, but we delete six in our analysis, as they describe methodical approaches or consist of self-references and words without economic meaning.

⁷ We approximate the research institutes information sets in terms of a short-term interest rate, the returns of the oil price (West Texas Intermediate), the returns of the real effective exchange rate, and the growth rate of industrial production (see Döpke and Fritsche 2006). Like Behrens, Pierdzioch, and Risse (2018a, 2018b), we take into account a forecast formation lag (that is, we assume that the research institutes use macroeconomic data for the month preceding the month in which a forecast is formed) and publication lags.

⁸ When we study pooled data, we reject strong as well as global strong forecast efficiency (at the 1% level) with the $GD{P}_{q}4$ forecasts being the only exception. The significant topics are the two investment topics for $GD{P}_{q2}$, government spending and taxes/social insurances for $CP{I}_{q2}$, as well as government spending and private/public-expenses for $CP{I}_{q4}$. For differences of tests results for individual institutes and pooled across institutes, see Döpke, Fritsche, and Siliverstovs (2010).

⁹ Estimating Equation (3)(3) $I_{t+h}=\alpha +\beta T_t+{\epsilon }_t,$(3) as a logit model yields qualitatively similar results (available upon request).

Additional information

Funding

This research was supported by the Deutsche Forschungsgemeinschaft [Project: Exploring the experience-expectation nexus in macroeconomic forecasting using computational textanalysis and machine learning; Project number: 275693836 / grant number: FR 2677/4-1).