Predicting recessions using trends in the yield spread

ABSTRACT

The yield spread, measured as the difference between long- and short-term interest rates, is widely regarded as one of the strongest predictors of economic recessions. In this paper, we propose an enhanced recession prediction model that incorporates trends in the value of the yield spread. We expect our model to generate stronger recession signals because a steadily declining value of the yield spread typically indicates growing pessimism associated with the reduced future business activity. We capture trends in the yield spread by considering both the level of the yield spread at a lag of 12 months as well as its value at each of the previous two quarters leading up to the forecast origin, and we evaluate its predictive abilities using both logit and artificial neural network models. Our results indicate that models incorporating information from the time series of the yield spread correctly predict future recession periods much better than models only considering the spread value as of the forecast origin. Furthermore, the results are strongest for our artificial neural network model and logistic regression model that includes interaction terms, which we confirm using both a blocked cross-validation technique as well as an expanding estimation window approach.

KEYWORDS:

• Forecasting
• classification
• logistic regression
• artificial neural networks
• cross-validation

JEL CLASSIFICATIONS:

• C45
• C52
• C53
• E43

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. The NBER business cycle dates offer the benefit of not relying on a specific indicator of economic activity and instead focusing on economic activity as a whole. The business cycle dating committee states that it ‘examines and compares the behavior of various measures of broad activity: real GDP measured on the product and income sides, economy-wide employment, and real income’, and may also consider additional indicators.

2. The two approaches naturally yield identical probability forecasts in the case of the logistic regression models and have minimal effect on our ANN model results. For instance, we computed out-of-sample ANN probability forecasts using both sets of inputs and found their correlation exceeded 0.98.

3. In unreported results we compute AUROC for our different models and find this metric generally ranks the models similarly to our other measures.

4. As with the logit model, LM, an equivalent representation in terms of yield spread changes can be made for our interaction model. Given our focus on forecast performance rather than interpreting the individual coefficients, we opt for this straightforward model representation.

5. Business cycle data available at http://www.nber.org/cycles/cyclesmain.html and U.S. Treasury interest rate data available at https://www.federalreserve.gov/releases/h15/.

6. Golosnoy and Hogrefe [11] notes that official NBER business cycle dates are announced with a considerable lag. For instance, the start date of the recession that officially began in December 2007 was not announced until December 2008. Our models attempt to identify these periods of expansion or contraction in economic activity one year before they occur, and it is even longer before they are officially defined.

7. The residual autocorrelation is small and statistically insignificant for all models well before 12 months; however, removing fewer months has minimal effect on the results. As noted in previous work in this area, the forecasting error is not usually a white noise process and may have temporal relationships; however, this does not create problems in our study since we do not make any statistical inference based on the forecasting errors.

8. We report additional statistics that do not involve a choice of cutoff for our expanding-window tests but only report sensitivity and specificity here in the interest of space.

9. We omit the figures for models L15 and L18 in the interest of space and to focus solely on the models with a one-year forecast horizon.

10. Most prior studies rely on a single performance measure. We choose to report a variety of measures, because each provides additional information and, as noted in [2], all individual measures have certain shortcomings.