Volatility-resilient cryptocurrency forecasting using deep temporal models and metaheuristic optimization

This paper presents a hybrid deep learning framework for cryptocurrency price forecasting that integrates complementary modeling, feature selection, and optimization techniques to address the volatility and nonstationarity of financial time series. The framework combines Convolutional Neural Networks for local pattern extraction with Peephole Long Short-Term Memory and Gated Recurrent Unit layers within an encoder–decoder architecture enhanced by temporal attention. Feature redundancy is reduced using a Genetic Algorithm–based selection strategy applied to common technical indicators, including moving averages, relative strength index, and Bollinger Bands. Model training employs a hybrid optimization scheme in which Bayesian Optimization performs global hyperparameter tuning, while Adam provides adaptive gradient-based updates. The framework is evaluated as a unified benchmarking pipeline on historical Bitcoin, Ethereum, and Litecoin price data under identical preprocessing and optimization settings. Experimental results demonstrate consistent improvements over baseline models, achieving RMSE values of 1049.69 for Bitcoin, 88.33 for Ethereum, and 3.47 for Litecoin, corresponding to relative error reductions of 35–60% across assets. A simulation-based trading analysis further validates the practical effectiveness of the proposed approach under realistic market conditions. Overall, the study shows that structured hybridization of deep learning architectures, combined with automated feature selection and optimization, enhances forecasting robustness in volatile cryptocurrency markets.