Regime Learning and Prediction on Time-series Data

点击翻译按钮获取中文摘要

This project investigates regime identification and characterization, and regime switching prediction in time series data, with applications in financial data analysis. A regime corresponds to a dynamic structure such as volatility, trends, patterns, dependence between financial assets. Regime switch analysis aims to detect such structures and their duration, and to forecast whether and how the structure changes from one regime to another. It is a key issue in many fields including climate dynamics, environmental ecology, financial economics, energy, meteorology, medicine, even tourism. As examples, portfolio managers of hedge funds could benefit from understanding the implications of regime switches in a financial ecosystem since the prediction of upcoming structural changes allows them to optimize the investment decisions, i.e. maximizing profits and reducing risks. Understanding regime switches in electricity consumption helps an electricity company forecast loads and manage the production or purchase of electricity. This project aims to discover cross-sectional market dynamics by investigating a new approach to regime identification and RS prediction. By approaching the problem from a data mining perspective, we will build a novel non-parametric/semi-parametric framework that eases the discovery of dynamics and features in the financial ecosystem. Our approach includes several innovative components. We search for and explore sequential patterns and trajectories in feature spaces to characterize regimes. We assess effects of outliers on regime changes. We search for and explore trajectories at the regime space in order to enable the use of non-linear machine learning models for RS prediction. We seek to identify causality rules and relationships between financial assets and their effectiveness in predicting important financial events, to provide insights on regime formation and duration and to leverage these insights for portfolio management. We design and develop new prediction models by exploring survival analysis, deep learning and pattern features.