Hybrid Optimal Time Series Modeling for Cryptocurrency Price Prediction: Feature Selection, Structure and Hyperparameter Optimization

Abstract

The prime aim of the research is to forecast the future value of Bitcoin which is commonly known as a pioneer of the Cryptocurrency market by constructing a hybrid structure over the time series. In this perspective, two separate hybrid structures were created by using Artificial Neural Network (ANN) together with Genetic Algorithm (GA) and Particle Swarm Optimization Algorithm (PSO). By using the hybrid structures created, both the network model and the hyperparameters in the network structure, together with the time intervals of the daily closing prices and how much data should be taken retrospectively, were optimized. Employing the created GAANN (DCP1) and PSO-ANN (DCP2) hybrid structures and the 721-day Bitcoin series, the goal of accurately predicting the values that Bitcoin will receive has been achieved. According to the comparative results obtained in line with the stated objectives and targets, it has been determined that the structure obtained with the DCP1 hybrid model has a success rate of 99% and 97.54% in training and validation, respectively. It should also, be underlined that the DCP1 model showed 47% better results than the DCP2 hybrid model. With the proposed hybrid structure, the network parameters and network model that should be used in the ANN network structure are optimized in order to obtain more efficient results in cryptocurrency price forecasting, while optimizing which input data should be used in terms of frequency and closing price to be chosen.