Multiscale forecasting and risk measurement in the crude oil market

Abstract

With the increasing trend of globalization and deregulation comes the increasing level of structural complexity in the crude oil market, which in turn leads to higher

uctuations in its price. Thus the current methodologies of predicting oil prices and measuring are being challenged as they offer only moderate levels of accuracy and reliability. To face these challenges, this thesis attempts to gain better understanding of the underlying market structure by integrating a recently revealed stylized fact, heterogeneous multiscale market structure, into the modeling process. Emerging multiscale signal processing techniques in engineering fields including wavelet analysis and Morphological Component Analysis (MCA) are used during the modeling process and construction of various algorithms from the following five aspects. Firstly a multiple bases based denoising algorithm is proposed for forecasting crude oil price movement, that incorporates both multiscale techniques and machine learning based ensemble techniques. Empirical studies in the crude oil market show that finer separation of data with distinct characteristics lead to better modeling accuracy and generalizability. The nonlinear ensemble algorithm also contributes to stabilization of estimates based on different parameters. Secondly a Morphological Component Analysis (MCA) based forecasting algorithms is proposed for crude oil markets. The sparsity based MCA algorithm is introduced to analyze the heterogeneous micro structure in crude oil price evolution. Empirical studies have shown that multiple bases better represent the market's micro structure and lead to improved forecasting accuracy. Thirdly a multiscale nonlinear ensemble algorithm is proposed for estimating Value at Risk (VaR). The proposed algorithms incorporates both multiscale techniques and machine learning based ensemble techniques. Empirical studies in the crude oil market have shown that the market has heterogeneous micro structure and superior performance improvement results from the finer modeling. Fourthly a MCA based VaR estimation algorithm is proposed for crude oil markets. The MCA model is introduced to extract transient events, which are modeled with time series models of different specifications. Then the estimated transient event risk level is used to adjust VaR estimated under normal market conditions. Empirical studies in the crude oil market have shown that incorporating transient events risks during the modeling process effectively improves reliability of the estimated VaR. Fifthly a multivariate denoising based algorithm is proposed for estimating crude oil Portfolio Value at Risk (PVaR). The proposed algorithms uses the multiscale denoising techniques to separate data and noises in the higher dimensional space to estimate crude oil PVaR. Previous studies have shown that the dependency among portfolio assets has heterogeneous multiscale structure. Effective separation and incorporation of this data features lead to improved performance. Work in this thesis provides alternative solutions for some extremely important practical problems in crude oil markets, which are of both theoretical and practical values. It would help business practitioners forecast price movement and track risk evolutions better. And it would extend and enrich the forecasting and risk measurement literature.