Predicting Recessions Using Boosting and Bayesian Model Averaging

Abstract

Economic recessions are costly, and are among other things associated with high unemployment rates, low wage growth, low investment spending and a higher number of bankruptcies. Whether the economy is in a recession or an expansion is important for economic policy decisions. In this thesis I compare the performance of two data-driven methods for predicting US recessions. The methods I use are called boosting and Bayesian model averaging (BMA). Boosting is a machine learning technique, while BMA is a Bayesian method. My main objective in this thesis is to predict recessions in the US. The dataset I use consist of 128 different economic and financial variables for the US from January 1959 to November 2018. I evaluate in-sample and out-of-sample performance of different boosting and BMA specifications for predicting recessions six months ahead. To assess the forecast accuracy, I calculate the receiver operating characteristic (ROC) curve. The forecast performance is evaluated by the integrated area under the ROC-curve (AUROC). The in-sample results for boosting and BMA show that both methods predict recessions well, with AUROC-values of above 0.9. The out-of-sample results are more mixed, where the AUROC-values are in most of the cases between 0.85 and 0.90. The out-of-sample AUROC-values that I obtain for BMA and boosting are in line with AUROC-values found in earlier published papers that use more traditional econometric models. Finally, although both BMA and boosting allow for including a large set of predictors, I find that only a few predictors are important for predicting recessions. Most of these variables are well-known for being informative about future recessions. Particularly, I find that different interest rate spreads are the most important predictors for US recessions.