Abstract
This study investigates the role of key economic variables (spanning sentiment, macro-economic, and financial indicators) in driving and predicting stock market bubbles in the United States. To achieve this objective, periods of stock market exuberance are first identified using the Generalised Supremum Augmented Dickey–Fuller (GSADF) test, which effectively detects multiple bubble episodes within a time series framework. Subsequently, the study employs a combination of machine learning feature-ranking techniques and advanced Bayesian econometric models to determine the most influential predictors of stock market bubbles and evaluate their predictive performance. The machine learning approaches applied, include Decision Tree, Neural Network, Gradient Boosting Tree, and Random Forest algorithms, which collectively enable a robust assessment of nonlinear relationships and variable importance.
Furthermore, a Bayesian Model Averaging (BMA) logistic model is utilised to incorporate model uncertainty in the identification and prediction of bubble episodes. The study finds that the Consumer Confidence Index Conference Board is the most important variable with regard to the key drivers of stock bubbles, highlighting the importance of sentiment in analysing stock bubbles. For predictive ability, the study recommends policymakers and investors to closely monitor consumer sentiments reflected in the Consumer Confidence Index Conference Board (CCI CB).