Price Trend Forecasting of Cryptocurrency Using Multiple Technical Indicators and SHAP

Abstract

Investment predicated on price trends stands as one of the most prevalent and efficacious approaches, hinging on its capacity to accurately discern the price trajectory for each asset. Such a pursuit poses itself as one of the most formidable challenges within the realm of investments. In this study, the application of machine learning models is employed, while simultaneously comparing their prognostic capabilities to evaluate their performance in forecasting cryptocurrency price trends. Additionally, the normalization technique and the Shapley Additive exPlanations (SHAP) feature selection method are employed to effectively augment the aptitude for projecting cryptocurrency price trends. The prediction period encompasses the time span from January 1, 2014, to December 31, 2021. The experimental findings reveal that the Support Vector Machine (SVM) outperforms other models such as K-Nearest Neighbors (KNN), Random Forest (RFC), Naïve Bayes, and Long short-term memory (LSTM) when forecasting periods extend 7, 15, and 30 days beyond the present, respectively. However, when the forecast horizon is extended to 90 days, the LSTM model exhibits the most optimal performance.