Roots from Trees: A Machine Learning Approach to Unit Root Detection

Abstract: In this paper we draw inspiration from the ensemble forecasting and model averaging literature and use a gradient descent boosting algorithm to exploit variation between test statistics used to determine if a series contains a unit root. The result is a pseudo-composite ML-based test for unit roots which is four to six percentage points more accurate than the next best traditional test. Through a train-validation framework this method allows for control over Type I error rates and the gains in power come with little variation in specificity (empirical size). Additionally, the proposed method is agnostic towards deterministic elements traditionally needed in the established testing environment and thus closes off an additional error path for unit root testing; that of model misspecification. We illustrate this new testing procedure by applying it to an established benchmark data set and examining the state-level hypothesis of unemployment hysteresis.