Built a deep reinforcement learning model for portfolio optimization. Based upon the AlphaPortfolio, constructed by Lin William Cong, et. al., my procedure takes financial features and generates a portfolio prescribing long and short positions to inputted asssets. The agent/critic algorithm optimizes a loss function that punishes excessive volatility and poor returns.
Developed a transformer encoder neural network model with an allocation head to give values between 0, representing no leverage entered into the market, and 2, representing 2x leverage into the S&P 500. This model first normalizes and embeds the input features into an embedding layer. It then applies positional encoding to provide the mechanism a sense of time ordering, which is then fed into the transformer encoder.
The goal of this project is to apply my physics experience to finance by modeling interconnected global markets as a network of coupled oscilllators (think pendulums together on a string). By analyzing key metrics such as the resulting coupling matrix K, one can uncover underlying coupling strengths between markets or assets, explore their synchronization, and perform stability analysis on the system as a whole. This is not meant to predict prices, but rather probe the dynamics of the system within a given period of time.
Interactive dashboard for real-time market analysis of the S&P 500 with technical indicators, statistics, and informative visualizations.