Multiscale Financial Signal Processing and Machine Learning

Abstract

Financial time series such as market indices and asset prices are shown to be driven by multiscale factors, ranging from long-term market regimes to rapid fluctuations. Multiscale analysis and signal processing not only reveal latent behaviors embedded in financial time series, but also help machine learning prediction tasks. In this thesis we focus on two different approaches tailored for daily and intraday financial time series respectively. In the first study, Hilbert-Huang transform is applied to daily prices and index values to reveal the underlying multiscale dynamics. In addition, a novel machine learning framework is proposed for identifying useful predictive features. An adaptive algorithm for highly nonstationary time series was introduced and applied to cryptocurrencies to show embedded structure and spectral properties. In the second study, we inspect the relations between statistical properties at different timescales, with the application to intraday high-frequency price data with noise. Functions describing the multiscale behaviors of volatility and correlation are defined and computed using empirical data. Models for high-frequency price processes are proposed and compared against empirical observations.