In many financial applications, it is important to classify time-series data without any latency while maintaining persistence in the identified states. The authors propose a greedy online classifier that contemporaneously determines which hidden state a new observation belongs to without the need to parse historical observations and without compromising persistence. Their classifier is based on the idea of clustering temporal features while explicitly penalizing jumps between states by a fixed-cost regularization term that can be calibrated to achieve a desired level of persistence. Through a series of return simulations, the authors show that in most settings their new classifier remarkably obtains a higher accuracy than the correctly specified maximum likelihood estimator. They illustrate that the new classifier is more robust to misspecification and yields state sequences that are significantly more persistent both in and out of sample. They demonstrate how classification accuracy can be further improved by including features that are based on intraday data. Finally, the authors apply the new classifier to estimate persistent states of the S&P 500 Index.
Subject classification (UKÄ)
- Probability Theory and Statistics