Efficient High-Frequency Cryptocurrency Forecasting: A Comparative Study of ARIMA and LSTM

This study examines high-frequency cryptocurrency price forecasting using two distinct time series models: the Autoregressive Integrated Moving Average (ARIMA) model and the Long Short-Term Memory (LSTM) neural network. The analysis is conducted using one-minute interval data for five major cryptocurrencies—Bitcoin (BTC), Ethereum (ETH), Binance Coin (BNB), Solana (SOL), and Ripple (XRP). Model performance is evaluated based on three standard error metrics: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). Empirical results indicate that ARIMA, despite its computational simplicity, achieves superior predictive accuracy for BTC, BNB, and SOL, with performance gains ranging from 17% to 100% compared to LSTM. For XRP and ETH—cryptocurrencies with lower price volatility—LSTM shows better performance in terms of relative prediction error (MAPE), suggesting its capability to capture nonlinear patterns in smoother data. This study also incorporates hyperparameter tuning for LSTM and explores model responsiveness, training time, and scalability for deployment in real-time trading systems. Findings reveal that while ARIMA is highly efficient and better suited for short-duration forecasts in resource-constrained environments, LSTM remains valuable for modeling more complex data with extended memory needs. These findings highlight the effectiveness of traditional statistical models in high-frequency trading contexts, especially when the data exhibit stationarity and strong short-term dependencies. The outcomes contribute valuable insights for the selection of forecasting models in algorithmic trading systems, real-time risk assessment, and data-driven decision-making in the field of digital finance.

Recommended citation: Sevani, N., Ardiansyah, A., & Sutrisno, H. (2025, July). Efficient High-Frequency Cryptocurrency Forecasting: A Comparative Study of ARIMA and LSTM. In 2025 International Conference on Data Science and Its Applications (ICoDSA) (pp. 751-757). IEEE.
Download Paper