Recession Signals: New Tools and Machine Learning Methods

Recession forecast for the U.S.

Jeffrey Frankel explains that the risk of a U.S. recession has risen, mainly due to trade tensions, market volatility, and unpredictable government policies. Key warning signs include an inverted yield curve, declining consumer and business confidence, and negative real-time GDP estimates. Professional forecasts and prediction markets now put the chance of a recession at about 55%. While some indicators, like unemployment, only confirm recessions after they start, others (like the Purchasing Managers Index) suggest weakness now. Frankel concludes that although the risk of recession is clearly higher than usual, forecasting remains uncertain and official confirmation always comes after the fact.

How to forecast a recession
By: Jeffrey Frankel – Harvard University

The “Michez” rule!

Pascal Michaillat and Emmanuel Saez develop a new, real-time rule (“Michez rule”) to detect US recessions by combining labor market data-specifically, job vacancies and unemployment rates. The goal is to improve the speed and reliability of recession detection compared to existing rules, notably the Sahm rule, which uses only unemployment data.

Construction of the Michez Rule:

Indicators used:
– Unemployment indicator: Measures the increase in the 3-month average unemployment rate above its 12-month low (as in the Sahm rule, but with minor adjustments for smoothing).
– Vacancy indicator: Measures the decrease in the 3-month average vacancy rate below its 12-month high (constructed analogously to the unemployment indicator).

Minimum indicator:
The core innovation is to take the minimum of the two indicators at each point in time. This “minimum indicator” becomes positive only when unemployment is rising and vacancies are falling, which is characteristic of recessions.

Threshold selection:
The recession threshold is set at 0.29 percentage points (pp) for the minimum indicator. A recession is detected when the minimum indicator crosses this threshold.  This threshold is the lowest value that avoids false positives (incorrectly signaling recessions) between 1960 and 2021.

Dual-threshold extension:
– A higher threshold (0.81pp) is introduced to quantify recession probability. Values between 0.29pp and 0.81pp suggest a probable recession; above 0.81pp, a recession is considered certain.  The probability of recession is computed linearly between these thresholds.

Data sources:
– Unemployment: Bureau of Labor Statistics (BLS) Current Population Survey.
– Vacancies: For 1960–2000, a help-wanted index (Barnichon, 2010); for 2001–2024, BLS Job Openings and Labor Turnover Survey (JOLTS).
– Historical data (1929–1959) are constructed from archival sources and proxies for vacancies.

Findings

• Performance (1960–2021):
  – Accuracy: The Michez rule detects all nine official US recessions in this period with no false positives or negatives-matching the Sahm rule’s accuracy for this era.
  – Timeliness: The Michez rule is significantly faster. Average detection delay: 1.2 months after recession start (vs. 2.7 months for the Sahm rule).  Maximum delay: 3 months (vs. 7 months for the Sahm rule). In all but one case (the 2008 Great Recession), the Michez rule detects recessions earlier than the Sahm rule.
• Historical Robustness (1929–1959):
  The Michez rule maintains perfect accuracy-detecting all six recessions with no false positives, even in periods of greater data volatility and less standardized measurement. The Sahm rule, by contrast, produces false positives in this era and cannot set a threshold that avoids both false positives and negatives.
• Current application (2022–2024):
  The Michez rule detected a recession in March 2024, when the minimum indicator crossed 0.29pp.  By December 2024, the minimum indicator was at 0.43pp, implying a 27% probability of recession (using the dual-threshold method). The Sahm rule also detected a recession in summer 2024 but gave a less persistent signal.

Has the Recession Started?
Authors: Pascal Michaillat, Emmanuel Saez
From: University of California, Santa Cruz, University of California, Berkeley

The Doombot!

Doombot, a custom machine-learning algorithm developed by Thomas Chalaux and Dave Turner at the OECD, consistently outperforms standard machine-learning models like LASSO and OCMT in predicting recession risks across 20 OECD countries, especially in out-of-sample tests. Its key benefits over other approaches are:

• Superior predictive performance: Despite imposing additional economic constraints, Doombot delivers the most reliable out-of-sample recession forecasts across multiple metrics and time horizons. It provided earlier and clearer warnings ahead of major events such as the 2008 Global Financial Crisis than other models.
• Avoids overfitting: Unlike LASSO, which often selects many variables and risks overfitting, Doombot uses a parsimonious set of typically fewer than three explanatory variables per equation, enhancing robustness and generalizability.
• Economic interpretability: Doombot imposes sign restrictions and smoothness constraints so that predictions are consistent with economic logic and avoid implausible, erratic jumps. This results in recession probability forecasts that are easier to interpret, explain, and communicate to policymakers-addressing the “black box” criticism common to many machine-learning models.
• Consistent and coherent narratives: By aligning variable selection and coefficient signs with economic theory, Doombot’s outputs provide a coherent economic story, making it easier to decompose and understand the drivers of recession risk in different countries and periods.
• Judgmental constraints improve performance: Contrary to concerns that restrictions might hinder machine learning, the authors finds that Doombot’s judgmental constraints actually improve out-of-sample predictive accuracy and narrative quality.

Doombot versus other machine-learning methods for evaluating recession risks in OECD countries
Authors: Thomas Chalaux, Dave Turner
From: OECD