A Comparative Analysis of ARIMA and LSTM Models for Bitcoin Price Prediction

A comparative study of two prediction techniques over Bitcoin price movements was carried out with ARIMA (AutoRegressive Integrated Moving Average), a statistical approach, and LSTM (Long Short-Term Memory), a specialized type of recurrent neural network (RNN) architecture, for predictive modelling of Bitcoin price movements. Yahoo Finance was employed to collect historical closing prices and obtain a comprehensive dataset that captures Bitcoin’s inherent volatility and non-stationary behaviour. Following data preprocessing, ARIMA parameters were identified based on stationarity tests and autocorrelation analyses, ensuring suitability for the dataset. Using a sequence-based approach, an LSTM model with two 50-unit layers was employed to capture complex temporal patterns, two Dropout layers were added to mitigate overfitting. Both models were trained and validated using metrics such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The experimental results offer insights into the strengths and limitations of each method. While ARIMA provides a robust baseline with interpretable parameters, LSTM demonstrates superior performance in capturing non-linear dependencies. Specifically, for daily forecasts, ARIMA achieves an MAE of 4,233.98 and an RMSE of 4,988.18, whereas LSTM reaches an MAE of 23,411.44 and an RMSE of 2,418.92. For minute-level forecasts, ARIMA yields an MAE of 248.07 and an RMSE of 306.55, while LSTM obtains an MAE of 1,582.51 and an RMSE of 59.00.