Under the supervision of Prof. Yuan Yao, I designed a trading algorithm using policy gradient to directly maximize the Sharpe ratio over a fixed period of time. The proposed algorithm performs significantly better than the Q learning baseline on a Bitcoin dataset.
Bitcoin Trading Agent with Deep Q-Learning Algorithms (Capstone 1)
I continued to explore the topic of my RIPS-HK 2018 group project and proposed some variants of deep Q learning trading algorithms.
Applying Q-Learning to Algorithmic Bitcoin Trading (RIPS-HK)
Chun Ho Chris Park, Matthew Thomas Sturm, Katherine Thai, Xinyi Linda Wang, Queenie Lee, Jonathan Yan. [report][repo]
Jun 2018 - Aug 2018
I participated in the Research in Industrial Projects for Students (RIPS-HK), sponsored by the HKUST Math department, IPAM at UCLA and RealAI. Our team designed and implemented several Q learning trading algorithms, all of which outperform the buy-and-hold strategy baseline.The poster of our project presented by Katherine won the “Outstanding Poster Award” at 2019 Joint Mathematics Meetings. [link]
NLP Application in Finance
Predicting Stock Volatility Using Domain Lexicon
Xinyi Wang, Yi Yang.
Sep 2018 - May 2019
Under the supervision of Prof. Yi Yang as a part-time student research assistant, I trained word embeddings on the financial documents with the incorporation of semantic information on different levels, then test the usefulness of the embeddings on the volatility prediction task. However, after comprehensive experiments, we concluded that our proposed method only outperforms the baselines by a very small margin.
Computer Vision in Biostatistics
Cell Counting by Adaptive Fully Convolutional Redundant Counting (Course project)
To enable fast domain transfer between different kinds of cells in the cell counting task, we propose to pre-train the network on a simple dataset, then freeze the domain agnostic parameters and only train the domain-specific parameter on a new dataset. Experiment results show that our proposed method significantly outperforms the training-from-scratch baselines.
