What does forecaster disagreement tell us about the state of the economy?

ABSTRACT

This article shows in a simple model that the part of uncertainty measured by forecaster disagreement rises in advance of and during recessions. Subsequently, it is tested using the Survey of Professional Forecasters in a dynamic probit model. It is shown that increases in disagreement help predict recessions in an out-of-sample context for the US.

KEYWORDS:

• Expectations
• SPF
• Uncertainty
• Dynamic Probit

JEL CLASSIFICATION:

• C22
• C52
• C53
• E17
• E37

Acknowledgments

We would like to thank an anonymous referee, Herman Stekler, James Foster, Jane Ryngaert, and Pavel Potuzak, as well as the participants in the IAES conference, the SAGE seminar, and the Futures Past: Economic Forecasting in the 20th and 21st Century workshop for their valuable comments and support.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Driver, Trapani, and Urga (2013) showed that disagreement can improve the prediction of underlying variables. Patton and Timmermann (2010) and Döpke and Fritsche (2006) both find that disagreement rises in recessions. For further information on disagreement and uncertainty, see Batchelor and Dua (1993), Bomberger (1996), Lahiri and Sheng (2010) and Ozturk and Sheng (2016).

² For real-time data, we use the second release which is released at the end of the second month of the quarter. This release has a slight informational advantage compared to the SPF forecast made earlier that month.

³ Results are similar if we use the median of the SPF, see Mboup and Wurtzel et al. (2018). We also explored models with CPI inflation and the unemployment rate. For inflation, we did not find a cyclical pattern in disagreement, consistent with the inconsistent cyclicality of inflation in this period. For the unemployment rate we found a weaker, but similar, pattern in disagreement to what we report for GDP growth.

⁴ Proan˜o and Theobald (2014) instead use three measures that are commonly used for continuous outcomes, namely root mean squared error (RMSE), mean absolute error (MAE), and the Theil coefficient. For robustness, we considered these measures as well and found that the results were broadly consistent with our findings based on ROC.

⁵ This analysis is pseudo out-of-sample in the sense that we are performing this analysis after the events occurred, but we estimated the model only based on data through Q4 of 1987 and 1995 respectively. Also note that if both the mean SPF forecast and disagreement are included, the out-of-sample results are indistinguishable from the ones of disagreement only.