A COMPARATIVE STUDY OF BIDIRECTIONAL LSTM-BASED PREDICTION FOR XRP, DOGE AND BNB

Abstract

Cryptocurrency markets are highly volatile, nonlinear, and influenced by rapidly changing market conditions, making accurate price forecasting a persistent challenge for researchers and practitioners. This study investigates the effectiveness of a Bidirectional Long Short-Term Memory (Bi-LSTM) neural network in predicting the daily closing prices of three widely traded digital assets: Ripple (XRP), Dogecoin (DOGE), and Binance Coin (BNB). Historical daily market data sourced from CoinMarketCap, covering the period from 1 January 2022 to 4 February 2025, were preprocessed using Microsoft Excel and modeled in Python with Pandas, NumPy, and TensorFlow/Keras. A 60-day rolling window of multivariate features: Open, High, Low, Close, Volume, and Market Capitalization was constructed to forecast the next day’s closing price. The model was trained using a 70:30 train–test split, 50 epochs, and a batch size of 32. Performance evaluation employed four key metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²). The Bi-LSTM achieved strong predictive accuracy across all three cryptocurrencies, with R² values of 0.82 for BNB, 0.98 for DOGE, and 0.96 for XRP. These results highlight the model’s ability to capture long-term dependencies and bidirectional temporal patterns inherent in cryptocurrency price movements, demonstrating its robustness for short-term forecasting in volatile markets. Overall, the findings suggest that Bi-LSTM models provide a reliable approach for cryptocurrency prediction tasks. The study concludes by offering recommendations and identifying opportunities for future work, including the integration of higher-frequency data, sentiment indicators, and more advanced deep-learning architectures.