Stacking Generalization via Machine Learning for Trend Detection in Financial Time Series

The task of understanding and modeling the dynamics of financial data has a significant practical value. In particular, it can help intercept trend inversion signals, providing an accurate future forecast that is important for asset allocation, investment planning, portfolio risk hedging and so on. Yet, the irregular fluctuations, chaotic dynamics and constantly changing patterns of financial data make time series modeling a challenging task in this domain. In this paper, we propose a classifier ensemble operator based on stacking generalization, which is applied to a pool of individual signals generated by a Poisson process-based model. The forecasting ability of the methodology is tested on a set of price time series. The results of the ensemble model application demonstrate the increased accuracy of prediction and a mitigated sensitivity of the model to parameters, outperforming the output of individual model components.