The Forecasting About Bitcoin and Other Digital Currency Markets: The Effects of Data Mining and Other Emerging Technologies

Abstract

We examine the bitcoin currency's predictability over prediction timeframes of 1 to 15 minutes. As a result, we evaluate a variety of machine learning methods and discover that, although all models perform better than a random predictor, Bidirectional recurrent neural networks and extreme Gradient Boosting Classifier (XGB Classifier) are particularly effective for the investigated prediction tasks. We employ a wide range of features, including technological, asset-based, interest-based, and block chain technology characteristics. Our findings indicate that technical characteristics continue to be the most important for the majority of techniques, followed by a few blockchain- and interest-based features. Furthermore, we see that reliability rises with increasing prediction horizons. Although a quantile-based long-short trading strategy may provide market return of up to 34% before transaction costs, the very brief holding durations result in negative yields after transaction expenses are incurred.