Statistical Prediction of the Cryptocurrency Bitcoin

Abstract

The purpose of this master thesis was to examine whether it’s possible to predict the future USD-returns of the cryptocurrency Bitcoin. This was conducted through the use of the statistical models Elastic Net, Random Forest, Gradient Boosting Machine and the Artificial Neural Network, where technical indicators were used as features – all within a walk-forward-optimization approach. The predictions from each of these models was to be used as inputs in their own respective trading strategy. The walk-forward-optimization approach allowed both the models’ hyperparameters and their corresponding trading strategy’s trading signals to adapt to potential changes in the BTC/USD market. Each trading strategy’s total return – conditional on a few market assumptions – was to be compared to a buy-and-hold strategy’s total return over the same period. Furthermore, the use of stacking was explored, in an attempt, to improve both the statistical results and the trading results. This implied constructing 4 new models, labeled meta-learners, based on the same algorithm-class as the above-mentioned models, labeled base-learners, but where the feature-space of each meta-learner contained the predictions of all four base-learners. The findings of this thesis suggest, that it’s somewhat possible to solve this prediction-problem, as all 4 base-learners achieved a lower mean-absolute-error than the 2 benchmark-models, mean- and zero-prediction. Additionally, the total return of each base-learner’s corresponding trading strategy produced a higher total return than the buy-and-hold strategy. It was also found, that the stacking approach was able to improve upon the statistical results, as out of all 8 models, 3 of the best 4 models were meta-learners and the best model was a metalearner. Finally, the use of stacking was able to improve the trading results as well, as 3 of the best trading strategies, conditional on the given market assumptions, contained predictions from meta-learners.