Predicting Bitcoin: a robust model for predicting Bitcoin price directions based on network influencers
Date
2016-10-01
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The ability to predict financial markets has tremendous potential to limit exposure
to risk and provide better assurances of annualized gains. In this thesis, a
model for predicting the future daily price of Bitcoin is proposed and evaluated
in comparison to that of a purely random model. Bitcoin is a novel digital currency
that relies on cryptography instead of a central authority to verify transactions.
Without a central authority, Bitcoin requires a complete list of all transactions
to be made public so that they can be verified by all users. This unique
feature of Bitcoin, where all transactions are public, is exploited by the model
to predict the future price directions based on the actions of Bitcoin users. The
daily activity of the markets, aggregate network features, and the actions of major
network influencers are all used as features for the predictive model. Where major
network influencers are defined as users that accumulate a disproportionate
amount of wealth within the Bitcoin network compared to others. The information
about the actions of all Bitcoin users are extracted from the blockchain and
stored in a relational database for ease of use. Two metrics were created to identify
the major network influencers based on the history of their actions recorded
on the blockchain. The first metric, based on the concept of an h-index, often
used in academia to rank authors by their citations, is used to rank users by the
amount of wealth they accumulate monthly. The second metric is based on the
optimization of multiple objectives, the maximum increase in wealth with the
least amount of activity using Pareto optimization. All of the major network influencers
identified were then used as features, in combination with aggregate
network features, and market data, to test and evaluate several predictive models.
The models created were based on non-linear equations, support vector machines,
decision trees, and XGBoost; all evaluated and compared using the same
data. The XGBoost model consistently proved to be much more accurate than all
other models and was used for the final set of experiments. The XGBoost model
was compared to that of a purely random Monte Carlo model using the entire
history of data for the period of 2013–2016. The first set of experiments were conducted
using various sizes of training and testing data, in each case the XGBoost
model had an accuracy 20% greater than that of the Monte Carlo model. For
the final experiments, the model was tested in a realistic scenario, predicting the
price direction for each future day, while also being re-trained using the results
of each new day. The XGBoost model achieved a much better performance in
comparison to the Monte Carlo model, which had approximately 50% accuracy,
whereas the XGBoost model had 70%–79% accuracy.
Description
Keywords
Bitcoin, Predictive model, Machine learning