Forecasting the Price of Bitcoin Using an Explainable CNN-LSTM Model

Artificial Intelligence (AI) significantly improves time series forecasting in the financial market, yet it is challenging to establish reliable real-world finance applications due to a lack of transparency and explainability. This paper prototypes an explainable CNN-LSTM model that combines the advantages of CNN and LSTM (Long and Short Term) to train and forecast the price of Bitcoin using a group of 11 determinants. By avoiding information loss and information superposition, it combines long-term context information and short-term feature information to obtain comprehensive and accurate feature representation. Experiments show that CNN-LSTM generally has higher accuracy than a single LSTM network when processing and predicting Bitcoin sequence data, as measured by a mean absolute percentage error (MAPE) of 2.39% and an accuracy of 89.54%. Additionally, the CNN-LSTM model explains that trading volume and prices (Low, High, Open) contribute to the price dynamics, while oil and Dow Jones Index (DJI) influence the price behavior at a low level. We argue that understanding these underlying explanatory determinants may increase the reliability of AI’s prediction in the cryptocurrency and general finance market.