Please use this identifier to cite or link to this item:
https://www.um.edu.mt/library/oar/handle/123456789/91843
Title: | Investigating the use of reinforcement learning in automated test case generation |
Authors: | Formosa, Matthew (2021) |
Keywords: | Machine learning Neural networks (Computer science) Deep learning (Machine learning) Stocks Exponentially weighted moving average |
Issue Date: | 2021 |
Citation: | Formosa, M. (2021). Investigating the use of reinforcement learning in automated test case generation (Master’s dissertation). |
Abstract: | The ability to identify price trends and patterns has always been a key interest for traders participating in the financial markets due to the associated potential gains. This is not a trivial task considering the complex nonlinear and noisy structure of the underlying price series. The problem gets more pronounced with higher frequency trading. As highlighted in a recent survey, less than 20% of the various literature they surveyed proposed a profitable trading system. Recent literature highlights the surge in deep learning popularity when applied to automating trading systems. A particular commonality in this research is that model features, typically technical indicators, are defined as functions at fixed time intervals. In our research, we focus on a recent event-based approach, namely features based on directional-change intrinsic time. We hypothesise that since market events are not bound by time, applying directional-change intrinsic time in conjunction with algorithmic trading models based on deep learning models outperform traditional batch, fixed time features, training approaches that suffer from the inability to adapt to changing market conditions. In our research we have explored whether popular technical indicator based features identified in literature can contribute in predicting directional-change intrinsic time events. Our results have shown that these features provide statistically significant prediction signals and also resulted in models that generated improved trading performance when compared to benchmark models utilising fixed time approaches. We have subsequently explored further model improvements by extending our models to an online learning approach as opposed to the traditional batch learning methods. Our results have shown that online machine learning models within the context of directional-change intrinsic time are able to significantly improve the risk-adjusted trading performance when compared to benchmarks from recent literature. |
Description: | M.Sc.(Melit.) |
URI: | https://www.um.edu.mt/library/oar/handle/123456789/91843 |
Appears in Collections: | Dissertations - FacICT - 2021 Dissertations - FacICTAI - 2021 |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
21MAIPT013.pdf Restricted Access | 3.09 MB | Adobe PDF | View/Open Request a copy |
Items in OAR@UM are protected by copyright, with all rights reserved, unless otherwise indicated.