Random forests track real-time recession risk across asset classes

The PICON model uses machine learning to infer recession probabilities over horizons ranging from one to four quarters by aggregating daily prices across equities, corporate credit, term premia, currencies and commodities.

Unlike traditional probit models that rely solely on the slope of the sovereign yield curve, the random forest framework captures non-linear interactions across asset classes.

In testing on US data since 1990, the model detected three of four recessions with implied probabilities exceeding 50 percent, while generating fewer false positives during the 2022–2024 yield curve inversion than the New York Fed benchmark.

During exogenous events like the 2020 pandemic, the signal remained low at 15 percent.

Using SHAP values to measure asset contributions, the authors show that equity and corporate bond valuations drive short-term recession signals, whereas yield curve slopes dominate at three to four quarters.

The model reveals that market-implied recession odds reached 43 percent in the euro area and 35 percent in the US in March 2022 following the invasion of Ukraine.

In April 2025, US tariff announcements lifted US recession odds to 21 percent, primarily through equity channels, while the August 2024 yen carry trade unwind registered as purely financial stress.