Expectation formation process in the European Central Bank Survey of Professional Forecasters

Abstract

This thesis studies the expectation formation mechanism of professional forecasters in the Euro area using surveyed forecast data of the European Central Bank’s Survey of Professional Forecasters (ECB SPF). The quarterly survey asks Euro area financial and non-financial institutions about their forecasts for the Euro area for the following macroeconomic variables: the harmonized index of consumer prices (HICP), the yearly growth rate of real gross domestic products and the unemployment rate. The survey tracks each individual forecaster with identifiers and has multiple overlapping forecast horizons making it possible to study how individual and aggregate forecasts evolve over time as new information emerges. The structure of the ECB SPF is utilized to construct direct micro level data estimates of information inattention parameter using the approach by Andrade and Le Bihan (2003). They calculated their estimates with the same dataset with a sample ranging from 2000 Q1 to 2012 Q4. In this thesis the estimates are updated with nine years of new data with the sample period being from 2000 Q1 to 2021 Q4. In addition, a regression approach introduced by Coibion and Gorodnichenko (2018) is exploited to study the predictability of forecast errors by forecast revisions. Both of the approaches aim, in essence, to study whether the dataset exhibits features rational inattention. The rational inattention models employed are the sticky information model (Mankiw and Reis 2002) and the noisy information model (Sims 2003 and Woodford 2002).

The main findings of Andrade and Le Bihan (2013) are reaffirmed by the empirical analysis with the extended sample with some differences. Forecasters disagree on their forecasts and furthermore, the disagreement increases if a revision is made. In other words, the disagreement among revising forecasters is on average greater than among those who do not revise their forecasts. There is also clear time variability in the disagreement that is explored and found that it could be driven by economic shocks by some simple bivariate regressions which is an intuitive result. When new information becomes available and forecasts are revised, the forecasters fail to systematically update their forecasts on all forecast horizons. They also fail to systematically update their forecasts across variables. The empirical estimates of the probability of updating forecasts for all three forecasted variables after a quarter of new information is below one indicating that there is stickiness of information present in the data. After one year of new information the attention is higher although still below one. The attention is also time varying and there are not many instances across the sample period where attention is complete and even fewer instances of such periods where the attention is complete at the same time for all three variables reflecting the finding that forecasters fail to update forecasts systematically across variables. In contrast to their results there are no signs of predictability with the regression tests in the forecast errors on inflation. Real GDP and unemployment rate do, however, exhibit signs of predictability by their past forecast errors and this is in line with previous findings.

In addition, the results of Coibion and Gorodnichenko (2018) about the predictability of forecast errors by forecast revisions did not arise in the data. The main difference to their work is the different dataset used with longer horizon between the instances used to construct the revisions (one year compared to one quarter in Coibion and Gorodnichenko). As the micro estimates point out that the probability of updating forecasts approaches one when the horizon approaches one year, thus it could render the stickiness effect that causes the predictability close to nothing. This may be, along with the different dataset, the driver behind the difference in results.

All in all, features of rational inattention characterizing the models explored in this thesis are found in the data. The novelty value of this thesis is to comprehensively review the literature around information inattention models and to extend the dataset with the same method by Andrade and LeBihan (2013) and to aply the approach of Coibion and Gorodnichenko (2018) to the ECB SPF dataset that has not been done to my knowledge before.