Implementasi Algoritma Long Short Term Memory (LSTM) Untuk Prediksi Harga Bitcoin

Abstract

The current study was taken forward in pursuance with the performance of potential effective technical indicators as embedded within an LSTM model used to predict daily Bitcoin prices. The data used was five years' worth of historical data for Bitcoin, sourced from Binance for the period of January 2020 to June 2025. Technical indicators have been calculated for raw prices along with others: RSI-14, EMA-12, EMA-26, and MACD. From a feature selection perspective, mutual information and pearson correlation were used to select the most predictive features. This has summed down to the final model only using the following six features for predictions: close, high, low, open, EMA_12, and EMA_26. The dataset is split temporally such that 70% is for training, 15% for validation, and 15% for testing. MinMaxScaler was fitted only on the training set to avoid data leakage. For the LSTM architecture, 64 units were used along with a Dense layer, 32 neurons, a dropout rate of 0.1, a learning rate set at 0.002, and a batch size of 16. The results show that the introduced improvements in the Model vastly outperformed the OHLCV-only baseline, to a significant extent: mean absolute error drops from $2,662.51 to $2,166.24; root mean square error reduces from $3,422.25 to $2,599.49; mean absolute percentage error reduces from 6.32 to 5.13%. These results further confirm that technical indicators encode valuable information about market momentum and trend transitions, especially remarkable during high volatility. The model can be a possibility but under constrained abrupt market change conditions; it does not get a grip.