A Study on Optimal Parameters of Long Short-Term Memory (LSTM) Networks for Cryptocurrency Price Forecasting: A Case Study of Bitcoin, Dogecoin, and Ethereum

Abstract

Digital currencies such as Bitcoin (BTC), Ethereum (ETH), and Dogecoin (DOGE) are increasingly popular, emphasizing the importance of price forecasting. This study employs a Long Short-Term Memory (LSTM) artificial neural network to predict digital currency prices. Historical price data were collected from Yahoo Finance, spanning from January 7, 2020 to January 5, 2025. The data were preprocessed using Min-Max Scaling, and a 60-day historical window was used to forecast the next day’s price. Subsequently, an LSTM model was constructed by tuning three parameters including, the number of LSTM units, the Dropout Rate, and the number of Dense Layer units and its performance was evaluated using the root mean square error (RMSE). The experimental results revealed that for BTC, the model with 60 LSTM units, a Dropout Rate of 0.1, and 2 Dense Layer  units achieved the lowest RMSE of 1967.35; for DOGE, the model with 70 LSTM units, a Dropout Rate of 0.1, and 1 Dense Layer  unit achieved the lowest RMSE of 0.01242; and for ETH, the model with 70 LSTM units, a Dropout Rate of 0.1, and 1 Dense Layer units achieved the lowest RMSE of 116.54. These findings demonstrate that LSTM networks can effectively forecast digital currency price trends, although forecasting accuracy may be impacted by market volatility.