| Abstract: |
| In this talk, we will explore the challenge of forecasting Bitcoin price movements over various horizons - 1, 7, 14, and 30 days - from two perspectives: computer science and trading. We evaluated three distinct models: Support Vector Machines (SVM), Random Forest (RF), and Multi-Layer Perceptron (MLP).
We begin with an overview of the current state of financial forecasting using machine learning, highlighting key findings from previous studies and the limitations they faced. The computer science segment will detail our rigorous approach to the problem, starting with the dataset construction and the features included. We will outline a data preparation pipeline for training the models, followed by a forecasting algorithm designed to train, evaluate, and optimize hyperparameters for the three models. Our findings will reveal that SVM exhibited superior predictive capabilities.
In the trading section, we will discuss how we leveraged the SVM forecasts to create a long-only trading strategy. This part will demonstrate that while the SVM performs well in theory, it can also be applied to develop a potentially profitable trading strategy in practice. Join us to discover how these insights contribute to the intersection of machine learning and trading in the cryptocurrency market. |
|