In this dissertation, we propose an online machine learning technique – named Artificial Continuous Prediction Market (ACPM) – to predict the value of a continuous variable by (i) integrating a set of data streams from heterogeneous sources with time varying compositions such as changing the quality of data streams, (ii) integrating the results of several analysis models for each data source when the most suitable model for a given data source is not known a priori, (iii) dynamically weighting the prediction of each analysis model and data source to form the system prediction.We adapt the concept of prediction market, motivated by their success in forecasting accurately the outcome of many events [Nikolova and Sami, 2007]. Our proposed model instantiates a sequence of prediction markets in which artificial agents play the role of market participants. Agents participate in the markets with the objective of increasing their own utility and hence indirectly cause the markets to aggregate their knowledge. Each market is run in a number of rounds in which agents have the opportunity to send their prediction and bet to the market. At the end of each round, the aggregated prediction of the crowd is announced to all agents, which provides a signal to agents about the private information of other agents so they can adjust their beliefs accordingly. Once the true value of the record is known, agents are rewarded according to accuracy of their prediction. Using this information, agents update their models and knowledge, with the aim of improving their performance in future markets.This thesis proposes two trading strategies to be utilised by agents when participating in a market. While the first one is a naive constant strategy, the second one is an adaptive strategy based on Q-Learning technique [Watkins, 1989].We evaluate the performance of our model in different situations using real-world and synthetic data sets. Our results suggest that ACPM: i) is either better or very close to the best performing agents, ii) is resilient to the addition of agents with low performance, iii) outperforms many well-known machine learning models, iv) is resilient to quality drop-out in the best performing agents, v) adapts to changes in quality of agents predictions.
|Date of Award||30 Jun 2016|
|Supervisor||Marina De Vos (Supervisor) & Julian Padget (Supervisor)|