Multi-step forecasts of complex dynamical systems using soft-computing tools, with application to crude oil returns

Abstract

It is well known that crude oil plays a vital role in economic development. However, crude oil prices are sensitive to a large number of exogenous factors (such as speculation and OPEC behaviour) which result in short-term volatility shocks. This in turn makes it very difficult to forecast the price movement even in the short-term. This thesis aims to build tools to determine the direction of forecast crude oil returns, multi-steps ahead. The goal is to exploit domain knowledge of the crude oil market dynamics, and incorporate them into a black-box model to improve forecast accuracy and to increase the forecasting horizon. This research is driven by inadequacies in current forecast methods, and their adverse economic impacts. Our investigation begins by running a battery of tests to understand the underlying structure of crude oil prices and returns. For non-linearity in the structure of these series, we use an established test for independence, the BDS test. The Fuzzy Classifier System for non-linearity (FCS) proposed by Kaboudan (1999) and a time-domain test for non-linearity introduced by Barnett and Wolff (2005) are also used. Finally, we estimate the Lyapunov exponents to establish the existence of chaotic dependence in crude oil prices and returns. Our tests consistently show that the dynamic forces driving crude oil prices and returns are non-linear, and possibly of low dimension. Moreover, the FCS test shows evidence of high noise levels, with smoothing or noise reduction being necessary for achieving improved forecast accuracy. We conclude that it is possible to forecast the crude oil price using non-linear models providing noise control measures are applied; the best hit rate achieved for out-of sample was 61%. In addition, we present a number of constraints on the objective function to act as a direct form of domain knowledge and to guide the learning process of the model.A further problem facing short-term (daily and weekly) crude oil price forecasting is that most of the fundamental variables, such as supply, demand, inventory and GDP, are recorded on monthly or quarterly bases. This leaves us with a limited number of potential explanatory variables. This process would benefit from the incorporation of additional information hints to aid the forecasting process. We show several methods to create and assimilate new time series to act as supplementary information in the learning process for neural networks. These methods include: (i) using nonfinancial data from the search index information from Google Insight for Search for inclusion within the soft-computing model, (ii) creating a time series from OPEC meeting announcements using dummy variables and wavelet analysis, and (iii) using technical analysis transformation as domain-specific knowledge. Our results show the effectiveness of these methods, with some caveats. Finally, we propose a novel multi-agent model for the crude oil market. The goal of this model is to generate hints that can be used to aid the training of traditional ANN. Therefore, we test whether the output of an artificial market can generate useful information to improve the learning process of traditional neural networks. The best hit rate we achieved using the forecast of these agents as additional input to ANN was 58%.This thesis contributes to the body of literature by narrowing the gap between three interrelated fields: (i) energy economics, (ii) time-series econometrics, and (iii) soft-computing closer together in one structure.