Harnessing the wisdom of the crowds for OECD recession predictions

A key challenge in predicting recessions is distinguishing which factors matter at different forecasting horizons and across different countries. This column applies machine learning methods to identify recession indicators across 20 OECD countries and eight quarterly horizons. An ensemble of probit models is one of the top-performing methods, harnessing the ‘wisdom of crowds’. In the first and second quarters of the forecast horizon, measures of current economic momentum are the most critical predictors of recession probability. Beyond these quarters, financial cycle variables such as credit and house prices are the most reliable long-range indicators of future downturns.

Macroeconomic forecasters have long struggled to accurately pinpoint the timing of future recessions. In response, an alternative approach using probabilistic models has gained traction. However, much of the existing recession modelling literature is constrained in its scope. Most studies are limited to an individual country, primarily focusing on economies such as the US, or cover only a small number of nations. Furthermore, research that does consider larger panels of countries often focuses on rarer, high-impact events like fiscal or financial crises, typically relying on models that pool countries and assume a uniform effect from a common set of explanatory variables across different economies.

The inherent complexity of predicting recessions means that existing studies often fail to clearly distinguish which explanatory variables matter at different forecasting horizons. This problem is compounded by the fact that the optimal set of predictors plausibly differs between countries and between prediction horizons. To achieve accurate forecasting, a model must be able to adapt its focus, shifting attention from near-term momentum to long-term structural risks.

The challenge can be understood by thinking of economic forecasting like meteorological work. Accurate prediction requires consulting both a daily weather forecast and a long-range climate model. The short-term forecast one or two quarters ahead (Q1-Q2), dominated by activity variables, tells you about immediate storms or clear skies right over your head. The long-range forecast (Q3-Q8), dominated by financial cycle variables like credit and house prices, measures the deep oceanic currents and atmospheric pressure systems that determine whether the entire season will be stormy or mild. Both are necessary, but they draw on fundamentally different sets of predictors.

This column aims to overcome these limitations by applying a broad and robust estimation framework to 20 OECD countries across eight consecutive quarterly horizons. Making use of machine learning methods, many combinations of explanatory variables are explored, rather than relying on prior judgement to select a few preferred explanatory variables. Thus, rather than using just labour-market variables (Michaillat 2025) or the yield curve slope (Mai et al. 2019), predictive performance is improved by using a much broader set of indicators. The research can then provide detailed, quarter-by-quarter insights into the main drivers of OECD recession risks.

The predictive contest and the wisdom of the crowds

Researchers from major institutions, including the IMF, ECB, and the Bank of England, have identified random forests (RF), or closely related methods, as the most consistently effective machine-learning method for identifying crisis episodes, often deemed superior to traditional probit/logit modelling (Bluwstein et al. 2023, Hellwig 2021, IMF 2021, Jarmulska 2020). However, our recent paper (Chalaux et al. 2025) challenges this prevailing view, demonstrating that a customised algorithm based on enhanced probit modelling can match, and in some key areas surpass, the performance of random forests when predicting recession episodes across 20 OECD countries.

The key to this revitalisation of probit modelling lies in embracing the concept of ensemble forecasting, the ‘wisdom of crowds’. Thus, an ensemble of probit models as well as pooled random forests emerged as the two top-performing methods in out-of-sample testing, providing the most insightful comparison regarding the drivers of recession risks.

The core predictors of OECD recessions: A shift in focus

By averaging the contributions of six categories of explanatory variables – economic activity, interest rates, inflation, share prices, credit, and house prices – across all 20 OECD countries over the full sample, clear patterns emerge regarding the drivers of recession risk. A fundamental finding is that the importance of predictors is highly dependent on the forecast horizon (Figure 1).

1. The immediate forecast: Activity variables dominate Q1 and Q2

The study finds a consensus among the top-performing models that activity variables – measures related to current economic momentum, such as GDP, industrial production, unemployment, and capacity utilisation – are significantly more critical predictors in the first and second quarters of the forecast horizon than in subsequent quarters.

Applying the meteorological metaphor: activity variables are the ‘daily weather forecast’, giving paramount importance to immediate business conditions. As the forecast horizon shrinks, variables reflecting current economic momentum naturally possess the strongest short-term predictive power. This dependency on current conditions is evident in the models’ decomposition of risk. However, the influence of these activity variables drops off sharply after Q2 in the probit model predictions, demonstrating the need for different indicators for longer forecasts.

2. The long-range climate model: Financial cycles rule beyond Q2

As the forecast horizon extends beyond the immediate quarters, the models shift their reliance, indicating that financial cycle variables dominate. This aligns with prior research highlighting the importance of deep, slowly evolving imbalances in predicting future crises (Borio et al. 2019). Thus, credit and house prices are identified as the crucial predictors for quarters Q3 through Q8, explaining on average more than 50-60% of the ensemble probit model recession risks.

Financial cycle variables are the ‘long-range climate model’, measuring the deep-seated structural pressures in the economy. The increasing importance of credit and house prices from the second quarter onwards confirms that financial imbalances, which take time to build, are the most reliable long-range indicators of future downturns.

Figure 1 Contributions to probit ensemble model recession probability predictions

Averaged across all 20 countries over the full sample, 1980-2024

Note: The contributions of each explanatory variable are evaluated according to a formula that ensures that the decomposition of the overall probability into its constituent drivers adds up to the total, despite the non-linearity of the underlying probit functions (see note to Figure 5 in Chalaux et al. 2015). These contributions are then averaged across many well-fitting equations at each quarterly horizon. The chart then sums all probabilities at each horizon separately and displays the relative importance of each category of variable to the probability sum at each horizon.

3. The role of global context and policy indicators

Beyond the core activity and financial categories, other factors contribute consistently to recession risks:

• International variables (e.g. OECD-wide or euro-area-wide financial aggregates and the real oil price) play a strong and increasing role in predicting recession risks (although these are combined with the corresponding domestic variables in Figure 1). Their contribution rises markedly, explaining between 40% and 50% of risks at horizons of four quarters and beyond, and tends to be even more important for more open and smaller economies. This suggests that global indicators frequently outperform domestic indicators in explaining country-specific recession probabilities.
• Interest rate developments explain about 15% of recession risks across all quarterly horizons for both models. This category includes measures like the slope of the yield curve – an established predictor, especially in the US – and changes in short-term rates to capture monetary policy actions.
• Inflation variables show increased importance from Q4 onwards, explaining around 10-15% of risks.
• Share price developments matter most in the second and third quarters, but with a declining importance beyond that.

Concluding remarks

This research underscores that accurate recession risk assessment across diverse OECD countries demands models that operate on multiple quarterly horizons. In particular, there is a fundamental shift in predictor relevance from immediate activity variables to longer-term financial cycle indicators.

Returning to the meteorological analogy, the key takeaway is that reliable economic guidance requires synthesising both views: the daily weather forecast of activity variables ensures you prepare for the immediate quarter, but it is the long-range climate model of financial cycle variables that truly predicts whether the economic season will be defined by persistent storms. Both are necessary, but they draw on fundamentally different sets of predictors.

Source : VOXeu