Explainable hybrid deep learning framework for cryptocurrency price forecasting using CNN-BiLSTM-TCN and SHAP analysis

Abstract

• Proposed an explainable hybrid deep learning framework (CNN–BiLSTM–TCN) for Bitcoin price forecasting. • Developed a multivariate forecasting pipeline using OHLC features and sliding-window sequence generation. • Achieved superior performance over standalone models with MAE = 0.004438, RMSE = 0.007075, and R 2 = 0.999341 . • Validated robustness using 5-fold time-series cross-validation with stable predictive accuracy. • Integrated SHAP-based explainability to quantify temporal feature contributions and enhance model transparency. Cryptocurrency markets are highly volatile, non-linear, and have complex time dependencies which in turn makes accurate forecasting a difficult problem. This paper presents a new hybrid deep learning architecture that uses Convolutional Neural Network (CNN), Bidirectional Long-Short-Term Memory (Bi-LSTM), and Temporal Convolutional Network (TCN) in order to combine the benefits of a recurrent model and a convolutional method. The model is secondary trained a multi-feature cryptocurrency dataset which does not obey log returns and incorporates temporal dependencies and augmentation on generalization by different architectural connections. To add transparency and interpretability to the predictions, we add SHAP (SHapley Additive exPlanations) analysis to the hybrid model and can attribution of features to time steps. Our solution gives the user an insight into how and why their models may be performing as they do with regard to showing which parts of their time series are most relevant and how they interact with each other, rather than focusing only on empirical results. Experimental findings indicate that the proposed hybrid model is much better in comparison with individual deep learning models in terms of RMSE, MAE, R-Square, and SHAP analysis proves that the model makes reasonable decisions that are consistent and explainable. The work offers an encouraging approach to explainable financial prediction: it has a high predictive performance and explains how predictions are made. It is possible to apply the approach to other fields of high-frequency time series outside the sphere of cryptocurrencies.