Accurately forecasting recessions remains a critical issue in macroeconomic analysis. This column introduces a new machine learning algorithm to predict recession risks for 20 OECD countries over quarterly horizons up to two years. The Doombot algorithm tests many combinations of financial and business cycle variables, but also imposes restrictions consistent with basic macroeconomic logic. The algorithm outperforms competing models and provides predictions easier to interpret and explain, for example when applied to the Global Crisis. The findings highlight the usefulness of judgement constraints in machine learning applications.
The challenge of accurately forecasting recessions remains a critical issue in macroeconomic analysis, as major forecasting errors often stem from an inability to anticipate economic downturns (An et al. 2018, Abreu 2011). Recognising the difficulty of pinpointing the precise timing of business cycle turning points, an alternative approach is to use probabilistic models to assess recession risk. One common method employs probit or logit models to estimate recession probabilities using various financial and economic indicators. However, most existing studies are limited to individual countries, primarily the US (e.g. Estrella and Mishkin 1996, Hamilton 2010, Pike and Vazquez-Grande 2019, Delle Monache et al. 2022, Summers and Domash 2022) or at best cover a small number of countries (e.g. Nyberg 2010, Fornari and Lemke 2010). Research on larger sets of countries has often focused on rare crisis events – like financial or currency crises – using models that assume uniform explanatory variables across different economies. This assumption is problematic for recession prediction, as advanced economies exhibit diverse macroeconomic and financial characteristics, including varying sensitivities to interest rates, trade openness, and the cyclicality of housing markets.
The absence of any published work (at least that we are aware of) assessing recession risks simultaneously across many countries and multiple horizons is due to the selection of the optimal set of predictors being challenging. This is not only because this choice can plausibly differ between countries, but also because it is likely to differ depending on the prediction horizon. Moreover, the functional forms of these relationships are not evident a priori, complicating the modelling process further.
Machine learning and economic judgment in recession risk prediction
In recent work (Chalaux and Turner 2024), we apply a range of machine learning methods to this selection problem for 20 OECD countries over a range of quarterly horizons up to two years. The competing machine learning methods include different parameterisations of the widely used least absolute shrinkage and selection operator (LASSO) and the more recent One Covariate at a time Multiple Testing (OCMT) estimators. These are compared with a ‘Doombot’ algorithm, custom-built by the OECD authors, which uses brute force to test many combinations of variables, but in doing so imposes numerous restrictions consistent with basic macroeconomic logic.
Across all the algorithms, the most frequently selected variables are financial variables, especially those relating to credit and house prices, but also equity prices and various measures of interest rates (such as the slope of the yield curve). Business cycle variables – survey measures of capacity utilisation, industrial production, GDP, and unemployment – are also selected, but more frequently at shorter horizons of only one or two quarters. Importantly, the models select not only domestic indicators but also international aggregates, reflecting the interconnected nature of global economies, consistent with earlier OECD studies (Hermansen and Röhn 2016).
Doombot has the best out-of-sample predictive performance
In evaluating the performance of competing probabilistic models that are predicting the occurrence of a particular event (here, a recession) it is customary to place emphasis on out-of-sample predictive performance. This is because if the events being predicted are relatively rare, then models are particularly vulnerable to ‘over-fitting’ on the training sample and then prove poor at predicting the occurrence of the next event in real time.
Accordingly, the competing algorithms are applied to data for 20 OECD countries with an emphasis on out-of-sample testing using a rolling origin, including windows covering the Global Crisis (GFC). The Doombot algorithm has the best scores averaged across countries among three different performance metrics at different horizons and is most often the top ranked algorithm among individual countries. Moreover, a close examination of the different vintages of out-of-sample predictions also shows that it provides a much better warning of the Global Crisis than other algorithms, including the second-placed LASSO algorithm.
Doombot is also better for ‘storytelling’
As well as being the best performing algorithm for predicting future recessions, the models selected by Doombot typically provide a more convincing economic narrative because:
- The number of explanatory variables is relatively parsimonious, having an average of about three explanatory variables per equation, which is much less than any of the LASSO variants.
- The signs on the explanatory variables are constrained in line with economic priors (and so are consistent across countries), whereas the signs estimated by the other algorithms are often contradictory.
- The term structure of recession probabilities is constrained to be relatively smooth, whereas the incidence of erratic and implausible spikes is more common in the predictions of the competing algorithms.
These features make Doombot’s predictions easier to interpret and explain. For example, in the case of risks to the US evaluated just ahead of the global crisis (Figure 1), recession probabilities steadily increase and remain high across most of the projection horizon. In the immediate quarters they are driven mainly by the inverted slope of the yield curve and falling share prices, but with a pronounced inflection in house prices dominating risks for quarters beyond that. The same vintage of predictions made for other countries reveals that nearly half of all other countries have predicted recession probabilities that are at least 50% for at least two consecutive quarters. For these countries, as well as the other countries where recession risks are elevated, house prices and/or credit developments play a prominent role, emphasising the general role of excess credit and house prices underlying the global crisis.
Figure 1 Contributions to predicted recession probabilities for the US ahead of the global crisis
Out-of-sample projections made with data available in early December 2007
Note: The chart shows an approximate decomposition of the recession probabilities into the contribution from each explanatory variable. The predictions are made with the Doombot algorithm using data available in early December 2007. The US was in recession from 2008 Q3 to 2009 Q2, corresponding to the shaded background area.
Implications for policy and future research
The Doombot algorithm demonstrates that incorporating economic judgement into machine learning models can yield more coherent narratives without sacrificing predictive accuracy. Indeed, the use of judgemental constraints appears to enhance out-of-sample performance. This finding challenges the notion that restrictions inherently limit a model’s effectiveness. By reducing the risk of overfitting and ensuring consistency with economic logic, Doombot provides a robust tool for policymakers to monitor and mitigate recession risks.
The success of Doombot raises important questions for future research: Can judgmental constraints similarly improve machine learning applications in other areas of macroeconomics? Exploring this could pave the way for more reliable and interpretable economic forecasts, including enhancing our ability to anticipate and respond to economic downturns.
Source : VOXeu
