Cryptocurrencies are highly volatile due to factors such as speculation, lack of regulation, and technological or political changes. The objective of this study is to multi-step forecasting of cryptocurrency fluctuations using recurrent neural networks and hybrid models. The applied methodology was based on four phases: Obtaining the dataset; Preprocessing; Implementation of the models (RNN, LSTM, GRU, CNN-RNN, CNN-LSTM and CNN-GRU) and Evaluation (MAE, MAPE, RMSE and R2). The result with the CNN-GRU model obtained the lowest MAE (0.01533) with 50 units, while the GRU with the same configuration achieved an R2 of 0.999994. Although the hybrid models showed advantages in certain metrics, the less complex models such as GRU showed a better balance between accuracy and computational efficiency. This study demonstrates that the LSTM and GRU-based models are effective in capturing the volatile nature of cryptocurrencies. Careful optimization of the hyperparameters and integration of the preprocessed data are key to improving predictive accuracy. The results demonstrate the practical utility of these models for analysts and investors and lay the groundwork for future research involving sentiment analysis and additional data.