Friday, April 22 2022
12:30pm - 2:00pm
Masters Presentation
A Survey of Deep Learning Architectures for Algorithmic Cryptocurrency Trading

Independent of one’s belief in cryptocurrency as an alternative to fiat currency, it is difficult to deny the enthusiasm of investors, both institutional and retail. Several properties of this novel asset set it apart and demand advanced methods of analysis and prediction. One area is in algorithmic trading, wherein a program extracts data, analyzes it, decides on an action, and executes that action on behalf of the investor. For this, many statistical techniques have been brought to bear on predicting how the market will behave in the immediate future for the decision-making process in the present. In this manuscript, we explore several statistical and artificial-intelligence methods applied to predicting price changes of the highest-market-valued cryptocurrency, Bitcoin, over a roughly three-and-a-half-year period of hourly data. Starting with creating a matrix of common market analysis techniques, known as Technical Indicators, and applying Principal Component Analysis to reduce our features, we then create trading agents with the following classical statistical, Artificial Neural Network, and Reinforcement Learning models: Exponential Smoothing, Seasonal Autoregressive Integrated Moving Average with eXogenous variables, eXtreme Gradient Boosting, Multi-Layer Perceptron, Convolutional Neural Network, Long-Short Term Memory Recurrent Neural Network, and two variations of Deep Reinforcement Learning. Additionally, we compare their performance with a random agent, three agents trading from three technical indicators themselves, and Buy-and-Hold. We find that the best overall algorithm for optimizing a portfolio consisting solely of Bitcoin is a Convolutional Neural Network-Based Twin Delayed Deep Deterministic Policy Gradient. This algorithm combines two neural networks, an actor capable of making decisions in a continuous action space and a critic capable of critiquing those actions in a continuous state space, allowing for more nuanced investing decisions in an uncertain market environment.
Speaker:Christopher Clark
Affiliation:
Location:SCB 4119


Download as iCalendar

Done