Search

Financial time-series are continuously generated by multiple sources, such as banks and corporations. These time-series are ordered sequences of data records that become available over time, imposing an order that must be considered when training models and making predictions. In addition, financial data represent, and are part of, highly complex systems that depend on and are generated by various factors such as financial policies and national economic growths. Thus, unlike traditional regression, predicting financial time-series requires consideration of both their sequential and interdependent nature. This thesis aims to address three related weaknesses of data-centralized and off-line machine learning methods: 1) Sequence Learning. Traditional approaches, developed to optimize per- formance on static data sets, restrict their flexibility to solve sequence pre- diction tasks in real-world applications. This means that, as traditional approaches are trained off-line, the profitable conditions of the trained models may disappear when they are tested in on-line environments; 2) Higher Order Statistics. The goal of dynamic modelling is to identify the dynamical system that produced a given input-output mapping. The mean square error (MSE), which is a second-order statistical measure, have been traditionally used as cost function when training adaptive systems. However, financial markets are complex and chaotic systems, meaning that second-order measures may be poor descriptors of optimality; 3) Distributed Learning. The movement of stock markets may affect the behaviour of stocks in other regions or countries. Traditional approaches do not directly consider inter-dependencies of the financial system, discarding interconnections and correlations that may represent important internal forces of the market. To address these issues, this work proposes a variety of kernel adaptive filtering approaches, which are data-driven methods for sequence learning that combine the convex optimization of linear adaptive filters and the universal approximation property of neural networks (NNs). The proposed approaches consider interconnections between stock markets, which reduces volatility and maximises returns while the robustness and simplicity of kernel adaptive filters are maintained. In particular, this thesis proposes: (1) a kernel adaptive filtering approach to support sequence prediction tasks in financial time-series; (2) an entropy-based cost function for kernel adaptive filtering to capture higher-order statistics of financial time-series; (3) a kernel adaptive filtering approach that captures internal forces of the market to improve profitability in real-time applications. The results show relatively low MSE values and higher Sharpe ratio when compared with recurrent NNs and autoregressive-based models. The proposed approaches, unlike NNs, do not need the whole training set to start learning the model, meaning that predictions are generated while the model is sequentially updated at the same time. In addition, when compared with kernel adaptive filtering methods, there is an improvement between 0.5% and 54% in terms of MSE, showing high tolerance to noisy and non-stationary conditions. The results of this thesis are in line with previous studies, suggesting that the United States market is more influenced by the European and not vice versa.

This thesis investigates decision-making in two-player imperfect information games against opponents whose actions can affect our rewards, and whose strategies may be based on memories of interaction, or may be changing, or both. The focus is on modelling these dynamic opponents, and using the models to learn high-reward strategies. The main contributions of this work are: 1. An approach to learn high-reward strategies in small simultaneous-move games against these opponents. This is done by using a model of the opponent learnt from sequence prediction, with (possibly discounted) rewards learnt from reinforcement learning, to lookahead using explicit tree search. Empirical results show that this gains higher average rewards per game than state-of-the-art reinforcement learning agents in three simultaneous-move games. They also show that several sequence prediction methods model these opponents effectively, supporting the idea of using them from areas such as data compression and string matching; 2. An online expectation-maximisation algorithm that infers an agent's hidden information based on its behaviour in imperfect information games; 3. An approach to learn high-reward strategies in medium-size sequential-move poker games against these opponents. This is done by using a model of the opponent learnt from sequence prediction, which needs its hidden information (inferred by the online expectation-maximisation algorithm), to train a state-of-the-art no-regret learning algorithm by simulating games between the algorithm and the model. Empirical results show that this improves the no-regret learning algorithm's rewards when playing against popular and state-of-the-art algorithms in two simplified poker games; 4. Demonstrating that several change detection methods can effectively model changing categorical distributions with experimental results comparing their accuracies to empirical distributions. These results also show that their models can be used to outperform state-of-the-art reinforcement learning agents in two simultaneous-move games. This supports the idea of modelling changing opponent strategies with change detection methods; 5. Experimental results for the self-play convergence to mixed strategy Nash equilibria of the empirical distributions of plays of sequence prediction and change detection methods. The results show that they converge faster, and in more cases for change detection, than fictitious play.

Protein X-ray crystallography images are usually studied for the purpose of identifying the three-dimensional structure of proteins at atomic resolution. These experiments produce hundreds and thousands of experimental results, leading to the need for automatic rather than manual analysis. The experiments are also time dependent and require monitoring over a long period of time, beginning from an hour of starting the experiment to a couple of months. The time gap between the imaged instances of these experiments also varies from frame to frame. This thesis introduces the use of difference images of an x-ray crystallography sequence in classifying the protein x-ray crystallography images, and exploits their various features along with the features of the original images. By further experiments using several combinations of difference images along with the original images, the significance of difference images in accurate classification of x-ray crystallography is established. At this stage, two independent feature sets for protein x-ray crystallography images are available. Based on these two feature sets, multi-view learning is employed for the purpose of classification of x-ray crystallography images, and random forests is shown to be a good base classifier for both views. During multi-view learning, different co-training methods are experimented with to determine the best possible classification result for protein x-ray crystallography images. After establishing by manual analysis that the same outcomes can be obtained via several different pathways in a crystallography time sequence, the results obtained from single frame classification experiments are used in the classification of protein x-ray crystallography sequences. Finally, the time varying nature of the x-ray crystallography experiments is further exploited together with the results of the previous experiments, to predict the final outcomes of crystallography experiments based on the first few frames. This provides the ability, for the first time, to determine whether to continue observing a sequence or to discard it as likely to be a non-crystallisation outcome.