The global currency market, which has millions of active participants and is the largest financial market in the world, has been greatly affected by the emergence of deep learning techniques. There is widespread recognition of deep learning’s effectiveness for price movement predictions, particularly in financial markets where its effectiveness has been shown. When compared to other machine learning methods, hybrid models like CNN-LSTM have shown to be the most effective and efficient. Nevertheless, there is a lack of thorough comparative research across deep learning hybrids, which makes it difficult to validate these claims empirically, especially in the diversified and extremely unpredictable FX market. To fill that void, this research builds a generic CNN-LSTM model to predict key currency pairs and compares it to another hybrid deep learning model. The objective is to evaluate and strengthen the reliability of deep learning techniques for FOREX forecasting. In order to maximize the generalizability of the model, this study mimics a categorization scenario using stringent cross-validation procedures. Important measures, including accuracy, precision, recall, and F1-score, were used to assess the models’ performance. The results show that there is a performance hierarchy, with CNN-LSTM, CNN-RNN. These findings highlight the need to evaluate deep learning models in different real-world contexts. In conclusion, this study highlights the difficulties of FOREX forecasting and the necessity for further evaluations of the predictive power of machine learning models in an ever-changing market.