This paper describes an algorithm, “DoomBot”, which selects parsimonious models to predict downturns over different quarterly horizons covering the ensuing two years for 20 OECD countries. The models are country- and horizon-specific and are automatically updated as the estimation sample period is extended, so facilitating out-of-sample evaluation of the algorithm. A limited combination of explanatory variables is chosen from a much larger pool of potential variables that include those that have been most useful in predicting downturns in previous OECD work. The most frequently selected variables are financial variables, especially those relating to credit and house prices, but also include equity prices and various measures of interest rates (such as the slope of the yield curve). Business cycle variables -- survey measure of capacity utilisation, industrial production, GDP and unemployment -- are also selected, but more frequently at very short horizons. The variables selected do not just relate to the domestic economy of the country being considered, but also international aggregates, consistent with findings from previous OECD work. The in-sample fit of the models is very good on standard performance metrics, although the out-of-sample performance is less impressive. The models do, however, provide a clear out-of-sample early warning of the Global Financial Crisis (GFC), especially when considered collectively, although they do generate ‘false alarms’ just ahead of the crisis. The models are less good at predicting the euro area crisis out-of-sample, but it is clear from the evolution of the choice of variables that the algorithm learns from this episode, for example through the more frequent selection of a variable measuring euro area sovereign bond spreads. The latest out-of-sample predictions made in mid-2023, suggest the probability of a downturn is at its greatest and most widespread since the GFC, with the largest contributions to such risks coming from house prices, interest rate developments (as measured by the slope of the yield curve and the rapidity of the change in short rates) and oil prices. On the other hand, warning signals from business cycle variables and equity prices, which are often good downturn predictors at short horizons, are conspicuously absent.
Doombot: a machine learning algorithm for predicting downturns in OECD countries
Working paper
OECD Economics Department Working Papers
Share
Facebook
Twitter
LinkedIn
Abstract
In the same series
-
Working paper20 September 2024
-
5 September 2024
-
5 September 2024