M.Sc. C.S. Proposal Defense: Lean Louiel A. Peria (Reinforcement learning system using temporal convolutional network and TimeGAN synthetic training data applied to portfolio optimization)
Dec. 16, 2021
Meeting ID: 831 3824 7855
Meeting Password: PeriaProp
Fredegusto Guido P. David, Ph.D., Chair
Adrian Roy L. Valdez, Ph.D., Adviser
Jaymar B. Soriano, D.Sc., Reader
Reinforcement learning is believed to be the most promising avenue of research in the field of algorithmic trading, seeing as how it leverages machine learning to perform the prediction of stock prices as well as decision-making for portfolio rebalancing at the same time. The objective of the study is to develop an RL trading system and apply it to historical stock price data from the Philippine Stock Exchange. A preliminary RL system has been implemented and backtests have already been performed however its performance is still unsatisfactory and requires more tuning. A training data augmentation module will be added to the RL system applying the TimeGAN technique of using a general adversarial network to generate synthetic historical stock price data for training. The TimeGAN method has already been applied to PSE data and the model appeared to produce decent benchmark scores. There is also a plan to apply a temporal convolutional network architecture to the policy network in hopes of improving performance.