Predicting Bitcoin: a robust model for predicting Bitcoin price directions based on network influencers

Abstract

The ability to predict financial markets has tremendous potential to limit exposure\nto risk and provide better assurances of annualized gains. In this thesis, a\nmodel for predicting the future daily price of Bitcoin is proposed and evaluated\nin comparison to that of a purely random model. Bitcoin is a novel digital currency\nthat relies on cryptography instead of a central authority to verify transactions.\nWithout a central authority, Bitcoin requires a complete list of all transactions\nto be made public so that they can be verified by all users. This unique\nfeature of Bitcoin, where all transactions are public, is exploited by the model\nto predict the future price directions based on the actions of Bitcoin users. The\ndaily activity of the markets, aggregate network features, and the actions of major\nnetwork influencers are all used as features for the predictive model. Where major\nnetwork influencers are defined as users that accumulate a disproportionate\namount of wealth within the Bitcoin network compared to others. The information\nabout the actions of all Bitcoin users are extracted from the blockchain and\nstored in a relational database for ease of use. Two metrics were created to identify\nthe major network influencers based on the history of their actions recorded\non the blockchain. The first metric, based on the concept of an h-index, often\nused in academia to rank authors by their citations, is used to rank users by the\namount of wealth they accumulate monthly. The second metric is based on the\noptimization of multiple objectives, the maximum increase in wealth with the\nleast amount of activity using Pareto optimization. All of the major network influencers\nidentified were then used as features, in combination with aggregate\nnetwork features, and market data, to test and evaluate several predictive models.\nThe models created were based on non-linear equations, support vector machines,\ndecision trees, and XGBoost; all evaluated and compared using the same\ndata. The XGBoost model consistently proved to be much more accurate than all\nother models and was used for the final set of experiments. The XGBoost model\nwas compared to that of a purely random Monte Carlo model using the entire\nhistory of data for the period of 2013???2016. The first set of experiments were conducted\nusing various sizes of training and testing data, in each case the XGBoost\nmodel had an accuracy 20% greater than that of the Monte Carlo model. For\nthe final experiments, the model was tested in a realistic scenario, predicting the\nprice direction for each future day, while also being re-trained using the results\nof each new day. The XGBoost model achieved a much better performance in\ncomparison to the Monte Carlo model, which had approximately 50% accuracy,\nwhereas the XGBoost model had 70%???79% accuracy.