Model-based Reinforcement Learning in Modified Lévy Jump-Diffusion Markov Decision Model and Its Financial Applications

This thesis intends to address an important cause of the 2007-2008 financial crisis by incorporating prediction on asset pricing jumps in asset pricing models, the non-normality of asset returns. Several different machine learning techniques, including the Unscented Kalman Filter and Approximate Planning are used, and an improvement in Approximate Planning is developed to improve algorithm time complexity with limited loss in optimality. We obtain significant result in predicting jumps with market sentiment memory extracted from Twitter. With the model, we develop a reinforcement learning module that achieves good performance and which captures over 60% of profitable periods in the market.