Bitcoin Price Prediction: Mixed Integer Quadratic Programming Versus Machine Learning Approaches

Abstract

Reliable Bitcoin price forecasts currently represent a challenging issue, due to the high volatility of this digital asset with respect to currencies in the Forex market. Since 2009 several models for Bitcoin price have been studied, based on neural networks, nonlinear optimization and regression approaches. More recently, Machine Learning paradigms have suggested novel ideas which provide successful guidelines. In particular, in this paper we start from considering the most recent performance of Bitcoin price, along with the history of its price, since they seem to partially invalidate well renowned regression models. This gives room to our Machine Learning and Mixed Integer Programming perspectives, since they seem to provide more reliable results. We remark that our outcomes are data–driven and do not need the fulfillment of standard assumptions required by regression–based approaches. Furthermore, considering the versatility of our approach, we allow the use of standard solvers for MIP optimization problems.