AbstractThis thesis presents an approach to support the identification and potential investigation of abnormal asset performance in traded securities, often popularised as ‘financial bubbles’. It uses an ensemble technique based on Case-based Reasoning (CBR) and Inverse Problems techniques (IPTs) that leverages capabilities of (CBR) in classification/prediction tasks especially in fuzzy domains and Strengths of (IPTs) in reconstructing ill-posed problems and identifying cause-effect relationships, to describe and model abnormal stock market fluctuations (often associated with asset bubbles) in time series datasets from historical stock market prices.
The thesis has developed and implemented a Machine Learning formative strategy called “IPCBR Framework” which is aimed to determine the causes of stock behaviour, rather than predicting future time series points in such fuzzy environment. By so doing, the research contributes to more robust strategies in investigating financial bubbles.
The IPCBR framework uses a rich set of past observations time series, and a geometric pattern description, and applies a combination of clustering techniques to derive a model that generalizes those patterns onto observations in the forward problem formulation. The derived result is then used as the input to the Inverse Problem formulation, the process of which is used to identify set of parameters that can statistically be associated with the occurrence of the observed patterns. The combined results is adapted to complete the CBR cycle.
This framework was implemented through the use of alliance machine learning methods.
The results of the implementation have shown that case retrieval accuracy can be achieved using a simple yet efficient approach that is based on assortment algorithms.
The thesis has demonstrated that, the Inverse solution can be successfully integrated into the CBR cycle, and that, given a target problem, the IPCBR framework provides a computationally inexpensive description of abnormal asset performance.
The research has contributed significantly to knowledge in various ways; a novel and more effective case representation of time series data has helped in solving the case retrieval issue since the time series do not perform well with the traditional attribute value representation in the CBR.
Secondly, the use of alliance machine learning methods has demonstrated case retrieval accuracy can be achieved using a simple yet efficient approach that is based on assortment algorithms. This in effect steer the course of producing the optimum combinations of classifiers through combining the strength of individual classifiers to arrive at an accurate classification. That, given the target problem, the IPCBR framework provides a computationally inexpensive description of abnormal asset performance.
Another major contribution to knowledge is in successful adaptation of the Sentiment analysis results as input to the CBR adaptation phase. Thais has proven to enhance CBR’s effectiveness in pattern matching.
This thesis has also demonstrated that the IPCBR framework can be successfully applied to the financial domain, which brings a novel perspective to the problem of asset bubbles investigation.
|Date of Award
|8 Oct 2021
|Stylianos Kapetanakis (Supervisor), George Samakovitis (Supervisor) & Miltos Miltos Petridis (Supervisor)