Forecasting Recessions Using an “At-Risk” Transformation to Identify Unusual Weakness in Indicators

Recession forecasts can significantly influence the decisions of households, investors, and businesses. And these forecasts provide policymakers with key information they can use to implement policies to minimize an economic downturn's impact on society. But forecasting recessions is challenging, and economists continually work to improve forecast accuracy and timeliness.

Standard recession forecasting models use a combination of economic and financial market indicators that economists deem useful in predicting future recessions. More recently, machine learning — which can accommodate the processing of extremely large data sets and the complex relationships between indicators — has opened new avenues for economists to forecast recessions.

For their paper, “At-Risk Transformation for U.S. Recession Prediction,” Senior Economic Advisor and Machine Learning Economist Minchul Shin of the Philadelphia Fed and Rahul Billakanti of Wayzata High School in Minnesota combined simple machine learning on/off “at-risk” signals of individual economic indicators to forecast U.S. recessions. An individual indicator is at-risk if it crosses its threshold, meaning it’s in a contractionary state compared with its historical behavior. Identifying these states, they write, is important because recessions are typically preceded by weakness in certain indicators. They also compare the performance of their binarization-of-predictors approach with the performance of standard forecasting models and a more complex machine learning model.

Other researchers have developed early-warning systems using single indicators such as the inverted yield curve¹ and the unemployment rate relative to a specific unemployment threshold.² But the paper by Shin and Billakanti is most aligned with a small-scale study that also uses a machine-learning-based binary signal transformation (which, interestingly, that study’s authors adapted from a pattern recognition algorithm of earthquake predictions);³ Shin and Billakanti advance this research by broadening the set of predictors and exploring multiple methods of aggregating the data.

Specifically, they used a database of monthly U.S. macroeconomic and financial indicators (122 indicators in total) from January 1960 to December 2024 to produce recession forecasts three, six, and 12 months ahead. But instead of using the raw historical data as in standard forecasting models and many other machine learning models, they transformed the data into signals, either 1 or 0, for each indicator series. In other words, the indicator variable takes on a value of 1 when it’s in an unusually weak position by historical standards and 0 when it’s in a typical state.⁴ Once they transformed the indicators into binary signals, their model operated like a standard forecasting model.

Their binarization-of-predictors approach produces recession forecasts that consistently outperform standard recession-forecasting models. Furthermore, their framework’s forecasts are competitive with or outperform forecasts from a more complex machine learning method (that is, XGBoost).⁵ Their binary-transformation approach “may appear to discard valuable information,” they write, but “it is well-suited for predicting rare events, such as U.S. recessions, where the relevant signal often lies in whether indicators cross into unusually adverse territory.”

Their framework performs particularly well at the onset of a recession. Catching a recession early is crucial from a practical perspective because it alerts economic participants of the potential impact on their jobs and finances, and it gives policymakers more lead time to implement monetary and fiscal policies.

When they use their binarized-predictors approach, they find that interest rates/spreads and the labor market were the dominant contributors, and of similar importance, in predicting the recessions of 1990, 2001, 2008, and 2020; money and credit, prices, and output and income also played meaningful roles (to varying degrees).⁶ The alternative forecast methods they analyze are much more dependent on interest rates/spreads in predicting recessions, with the other sectors contributing only marginally. In short, when they use their binarized predictors, a more diverse set of sectors informs recession forecasts.

Overall, the authors find that applying an at-risk transformation using simple binarization of predictors “is a powerful tool for recession forecasting.” Their approach has the added advantage of being easy to implement and computationally inexpensive, thus offering a practical benchmark for academic research and applied forecasting of rare events like recessions. Future work, they suggest, could test the effectiveness of their binarized-predictors method over longer data periods or apply it to comparable data sets from other countries.