Master of Science in Business Analytics
Master of Science in Business Analytics
Home
About
Program
Courses
Faculty
Application
Assistantship
Events
Projects
Resources
FAQ
Projects
2021-2022
Cryptocurrency Forecasting Using Machine Learning Models
Student Team
Lorena Andreas, Sonja Chill, Mustafa Khan, Raisa Silalahi, Phoebe Tran
Faculty Advisors
Dr. Mohammad Salehan, Dr. Mohamed Gomma
Purpose
Since its invention in 2009 Bitcoin has become an increasingly popular but highly volatile investment tool. We undertook this study in order to identify the algorithm most accurate at predicting the future price of Bitcoin in order to enable potential investment advisors to better guide investors.
Study Design/Methodology/Approach
We collected data from Kaggle.com, then developed and compared six separate time series prediction models: Autoregressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), Deep Autoregressive (AR), XGBoost, Linear Regression and Random Forest. The models were created using Python programming language on the Amazon Web Services platform. Models were trained using high-frequency market data of Bitcoin supplied by Kaggle.com. For comparison each model predicted the price of Bitcoin 15 minutes into the future with Mean Squared Error and Mean Percent Error used to compare model accuracy.
Findings
Price-related features are highly correlated. Using fewer features can significantly improve processing speed. The price pattern is also changing dynamically. Historical data from years or even months ago can be noise in predicting short-term prices. Our study found that the most accurate model for predicting the price of Bitcoin is the LSTM model trained on one month data.
Originality/Value
Our model achieves faster output and good performance using less data input compared with traditional studies using all the price-related features and all the historical data. It is important for our clients to quickly and accurately forecast short-term prices in real business.
Practical Implications
We recommend clients continue researching this topic to develop a more robust and accurate models. This project focuses on technical analysis. Trading strategies and social sentiment analysis using machine learning are beyond the scope. The models we train here still have limited value in terms of predictions.
Keywords
Cryptocurrency, Bitcoin, Forecasting, Time-Series, Machine Learning, Investing.