Forecasting the Unpredictable: Deep Neural Networks in the Volatile World of Cryptocurrencies

Cryptocurrency markets are known for their extreme volatility and nonstationary behavior, posing persistent challenges for accurate price forecasting. This study investigates the predictive performance of four deep learning architectures-Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and Transformer-for forecasting the HL2 price (average of high and low) of Bitcoin, Ethereum, and XRP across multiple temporal granularities (15-minute, I-hour, 4-hour, and 1-day intervals). Models are trained on historical candlestick data from Binance and evaluated using a rolling window approach over 100 consecutive forecast points. This setup enables a robust time-aware performance assessment. Evaluation metrics such as <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R^{2}$</tex>, MAE, RMSE, and MAPE are used to quantify predictive accuracy. Results show that the LSTM model consistently outperforms the others across all assets and intervals. The findings highlight the importance of temporal modeling and appropriate architecture selection in cryptocurrency forecasting. This comparative study contributes a systematic benchmark for deep learning-based forecasting using price-derived features across diverse market conditions.