Title page
Contents
Abstract/Resume 4
1. Introduction and summary 7
2. The broad framework 9
2.1. Country and time period coverage 10
2.2. Definition of a downturn 10
2.3. The potential set of variables explaining a downturn 11
2.4. Treatment of the pandemic period 14
3. The algorithm 14
3.1. Rules for model selection 14
3.2. A summary of the explanatory variables selected 15
3.3. A modified prediction rule when the economy is already in a downturn 17
4. Assessing performance 19
4.1. Comparison of Doombot with naïve forecasts and the Economic Outlook 20
4.2. Comparison of Doombot out-of-sample and in-sample performance 21
4.3. Out-of-sample performance in predicting the Global Financial Crisis 23
4.4. Out-of-sample performance in predicting the euro area crisis 24
5. Downturn risk predictions made in mid-2023 25
References 27
Annex A. Latest equations fitted values ("in-sample" forecasts) 29
Annex B. Recursive quarterly forecasts ("out-of-sample" forecasts) 40
Annex C. Country details of latest Doombot equations and predictions 43
Table 1. Downturn episodes since 1980 11
Table 2. Explanatory variables used to explain downturns 12
Table 3. The AUROC score across countries 22
Figure 1. Downturns are synchronised across countries 11
Figure 2. Selection of explanatory variables by country 16
Figure 3. Selection of explanatory variables by forecast horizon 17
Figure 4. The Poisson distribution as an approximation of the length of a downturn 18
Figure 5. The conditional probability distribution of the length of a downturn 19
Figure 6/Figure 4. F-score by forecast horizon 21
Figure 7/Figure 5. Distribution of out-of-sample downturn probabilities for 20 OECD countries 23
Figure 8/Figure 6. Comparison of current downturn probabilities with pre-GFC forecasts 25
Boxes
Box 1. Evaluating the performance of binary classification models 20
Annex Tables
Table A C.1. Variable and functional form notations 44