Exploring the experience-expectation nexus in macroeconomic forecasting using computational text analysis and machine learning

Prof. Dr. Jörg Döpke, Merseburg (Principal Investigator)

Prof. Dr. Ulrich Fritsche, Hamburg (Principal Investigator)

Prof. Dr. Christian Pierdzioch, Hamburg (Principal Investigator)

Dr. Sami Diaf, Hamburg

Tim Köhler, Merseburg


Based on the findings of the first project period and using the quantitative and qualitative data collected so far in the first work package we will use methods from corpus linguistics (sentiment analysis, topic models) to investigate the experience-expectation nexus and the role of ideology for the production of macroeconomic forecasts. The project is organized in three work packages.

The first package aims for an analysis of “ideological” positions of forecasters and the sequential processing of information and the evolution of topics describing business cycle expectations. So far, only few papers investigated the possibility of systematic biases of forecasts due to institutional affiliations or theoretical positions / ideological beliefs. Computational text analysis is a promising tool to measure ideological stances and topical representation of expectations.  Therefore, the package will use these methods to measure the possible a priori beliefs of institutions involved in forecasting and economic policy advice. Research in this area has a possible link to other fields and disciplines. There is a growing strand in the political science literature dealing with the measurement of political agendas and ideological positions using computer-based (corpus linguistic) methods

In the second work package, we investigate aspects of forecast optimality using non-linear methods and machine learning tools. For example, one problem with the traditional approach to test forecast optimality is that it neglects a potential nonlinear dependence of forecasts errors on its possible predictors. Such a nonlinearity can assume many different functional forms. For example, a particularly simple nonlinearity arises if the intercept of the orthogonality regression assumes different values depending on whether the nominal short-term interest rate is smaller (or even close to its zero-lower bound) than some threshold. In a similar vein, the traditional orthogonality regression model also neglects potential interaction effects between the predictors. For example, the shift in the intercept of the modified orthogonality regression model may occur only when the predictor variable assumes a value that is smaller than some threshold and, at the same time, another component is larger than some other threshold.

In the third work package, we will analyse additional insights into the quality of forecasts that might be obtained from quantification of textual or qualitative data. In particular, one might ask whether numbers and the forecasting story are always consistent. In a similar vein, the consistency of traditional forecast evaluation based on quantitative forecasts and an evaluation based on the forecast story is not certain. Therefore, based on recent approaches like the so-called “foredicton” analysis, we will test econometrically, whether a certain forecast story fits important aspects of the data.



Döpke, Jörg / Fritsche Ulrich / Waldhof, Gabi: Never Change a Losing Horse? On Adaptions in German Forecasting after the Great Financial Crisis, in: Fritsche, Ulrich / Köster, Roman /Lenel, Laetitia (eds.), Futures Past. Economic Forcasting in the 20th and 21st Century, Berlin et al.: Peter Lang 2020, pp. 191-218.

Fritsche, Ulrich / Köster, Roman /Lenel, Laetitia (eds.), Futures Past. Economic Forcasting in the 20th and 21st Century, Berlin et al.: Peter Lang 2020.