**7. Conclusions**

In this work, we explored the adoption of ensemble learning strategies with advanced deep learning models for forecasting cryptocurrency price and movement, which constitutes the main contribution of this research. The proposed ensemble models utilize state-of-the-art deep learning models as component learners, which are based on combinations of LSTM, BiLSTM and convolutional layers. An extensive and detailed experimental analysis was performed considering both classification and regression performance evaluation of averaging, bagging, and stacking ensemble strategies. Furthermore, the reliability and the efficiency of the predictions of each ensemble model was studied by examining for autocorrelation of the residuals.

Our numerical experiments revealed that ensemble learning and deep learning may efficiently be adapted to develop strong, stable, and reliable forecasting models. It is worth mentioning that due to the sensitivity of various hyper-parameters of the proposed ensemble models and their high complexity, it is possible that their prediction ability could be further improved by performing additional optimized configuration and mostly feature engineering. Nevertheless, in many real-world applications, the selection of the base learner as well as the specification of their number in an ensemble

strategy constitute a significant choice in terms of prediction accuracy, reliability, and computation time/cost. Actually, this fact acts as a limitation of our approach. The incorporation of deep learning models (which are by nature computational inefficient) in an ensemble learning approach, would lead the total training and prediction computation time to be considerably increased. Clearly, such an ensemble model would be inefficient on real-time and dynamic applications tasks with high-frequency inputs/outputs, compared to a single model. However, on low-frequency applications when the objective is the accuracy and reliability, such a model could significantly shine.

Our future work is concentrated on the development of an accurate and reliable decision support system for cryptocurrency forecasting enhanced with new performance metrics based on profits and returns. Additionally, an interesting idea which is worth investigating in the future is that in certain times of global instability, we experience a significant number of outliers in the prices of all cryptocurrencies. To address this problem an intelligent system might be developed based on an anomaly detection framework, utilizing unsupervised algorithms in order to "catch" outliers or other rare signals which could indicate cryptocurrency instability.

**Author Contributions:** Supervision, P.P.; Validation, E.P.; Writing—review & editing, I.E.L. and S.S. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Conflicts of Interest:** The authors declare no conflict of interest.

### **List of Acronyms and Abbreviations**

