ANALISIS MODEL LONG SHORT-TERM MEMORY (LSTM) DAN GATED RECURRENT UNIT (GRU) UNTUK PREDIKSI HARGA BITCOIN BERDASARKAN DATA CLOSING PRICE, VOLUME DAN INDIKATOR TEKNIKAL

Abstract

Blockchain technology has turned Bitcoin into one of the most heavily traded digital assets worldwide, but its sharp price swings make investment decisions difficult and call for a model that can capture complex, non-linear behavior. This study compares two Recurrent Neural Network (RNN) variants, Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), for Bitcoin price prediction using a combined input of closing price, trading volume, and three technical indicators, Moving Average (MA), Relative Strength Index (RSI), and Moving Average Convergence Divergence (MACD). The workflow consists of gathering daily Bitcoin price data, deriving the technical indicators, preprocessing the resulting dataset, training the LSTM and GRU networks, and scoring each model with Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). On the held-out test set, GRU obtained an RMSE of $12,194.34, an MAE of $10,125.90, a MAPE of 10.0629%, and 89.9371% accuracy, whereas LSTM obtained an RMSE of $26,401.57, an MAE of $23,476.99, a MAPE of 23.9027%, and 76.0973% accuracy. Both networks were able to track the broad price trend, yet GRU clearly outperformed LSTM under this multivariate setup, a gap attributed to its leaner gating structure and more efficient training. These findings position the resulting model as a possible building block for decision-support tools or automated-trading systems in cryptocurrency markets, adding to the growing body of deep-learning approaches for forecasting dynamic digital financial markets.