CONVOLUTIONAL NEURAL NETWORKS AND GRADIENT BOOSTING COMPARATIVE ANALYSIS FOR MULTI-HORIZON CLASSIFICATION OF CRYPTOCURRENCY STREAM DATA

Ivan Khamar

Abstract


Introduction. Cryptocurrency price direction classification is a binary classification task for time series with low composite signal/sum. Image encoding methods (GASF, MTF, Recurrence Plot) transform one-dimensional time series into two-dimensional images for processing by CNNs. However, a systematic comparison of the performance of such approaches with classical sequential and tabular methods for different data scales and forecasting horizons is lacking.

Materials and methods. A streaming pipeline based on Apache Kafka is proposed for collecting and processing OHLCV data of cryptocurrency trading pairs from the Binance exchange. Three series of experiments of different scales are conducted: small (46 pairs, 3 months, 10 948 samples), medium (91 pairs, 6 months, 46 355) and large (147 pairs, 12 months, 139 887). For each scale, six CNN architectures are compared – five with 2D image coding (SimpleCNN, HybridCNN, HybridResCNN, EfficientNet-B0, MobileNetV2) and one 1D control (CNN1D-Hybrid) – with gradient boosting (LightGBM) on tabular features. Classification was performed for four forecast horizons (1, 3, 6, and 12 intervals of 8 hours). Statistical significance was assessed using the McNemar criterion.

Results and Discussion. At a small scale, the 1D control model achieves the highest Matthews correlation coefficient (MCC) = +0.21 at horizon h = 1, while the best 2D image-based result is only MCC = +0.11. At medium scale, all CNN models exhibit MCC ≈ 0, while LightGBM consistently dominates with MCC up to +0.09 at horizons h = 6 and h = 12. At large scale, no model exceeds MCC = +0.07, indicating saturation of the predictive signal with increasing data heterogeneity. 2D image-based coding does not provide a consistent advantage over 1D sequential and tabular approaches at any scale.

Conclusion. The transformation of time series into two-dimensional images for convolutional neural networks does not provide a sustainable advantage over tabular gradient boosting and 1D convolutional networks in classifying cryptocurrency price direction. Naive scaling by increasing the number of trading pairs and time window worsens results of all model classes due to increasing heterogeneity of market regimes. These results indicate a need for regime-adaptive and cluster approaches instead of building a single global model.

Keywords: data classification; time series imaging; convolutional neural networks; gradient boosting; data streaming; Apache Kafka.


Full Text:

PDF

References


[1] Fang, F., Ventre, C., Basios, M., Kanthan, L., Martinez-Rego, D., Wu, F., & Li, L. (2022). Cryptocurrency trading: A comprehensive survey. Financial Innovation, 8(1), 13. https://doi.org/10.1186/s40854-021-00321-6

[2] Kreps, J., Narkhede, N., & Rao, J. (2011). Kafka: A distributed messaging system for log processing. Proceedings of the NetDB Workshop, 1–7.

[3] Wang, Z., & Oates, T. (2015). Imaging time-series to improve classification and imputation. Proceedings of the 24th International Conference on Artificial Intelligence (IJCAI), 3939–3945. https://doi.org/10.48550/arXiv.1506.00327

[4] Eckmann, J.-P., Kamphorst, S. O., & Ruelle, D. (1987). Recurrence plots of dynamical systems. Europhysics Letters, 4(9), 973–977.
https://doi.org/10.1209/0295-5075/4/9/004

[5] Hatami, N., Gavet, Y., & Debayle, J. (2018). Classification of time-series images using deep convolutional neural networks. Proceedings of the 10th International Conference on Machine Vision (ICMV), 10696, 106960Y. https://doi.org/10.1117/12.2309486

[6] Barra, S., Carta, S. M., Corriga, A., Podda, A. S., & Recupero, D. R. (2020). Deep learning and time series-to-image encoding for financial forecasting. IEEE/CAA Journal of Automatica Sinica, 7(3), 683–692. https://doi.org/10.1109/JAS.2020.1003132

[7] Chen, W., Xu, H., Jia, L., & Gao, Y. (2021). Machine learning model for Bitcoin exchange rate prediction using economic and technology determinants. International Journal of Forecasting, 37(1), 28–43. https://doi.org/10.1016/j.ijforecast.2020.02.008

[8] Jaquart, P., Dann, D., & Weinhardt, C. (2021). Short-term bitcoin market prediction via machine learning. Journal of Finance and Data Science, 7, 45–66. https://doi.org/10.1016/j.jfds.2021.03.001

[9] Qureshi, S. M., Saeed, A., Ahmad, F., Khattak, A. R., Almotiri, S. H., Al Ghamdi, M. A., & Rukh, M. S. (2025). Evaluating machine learning models for predictive accuracy in cryptocurrency price forecasting. PeerJ Computer Science, 11, e2626. https://doi.org/10.7717/peerj-cs.2626

[10] Khamar, I. (2026). CNN and gradient boosting comparative analysis for multi-horizon classification of cryptocurrency stream data [Kaggle Notebook]. Kaggle. https://www.kaggle.com/code/woodborder/v6-compvision-classification

[11] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770–778. https://doi.org/10.1109/CVPR.2016.90

[12] Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking model scaling for convolutional neural networks. Proceedings of the 36th International Conference on Machine Learning (ICML), 6105–6114. https://doi.org/10.48550/arXiv.1905.11946

[13] Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L.-C. (2018). MobileNetV2: Inverted residuals and linear bottlenecks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 4510–4520. https://doi.org/10.1109/CVPR.2018.00474

[14] Chicco, D., & Jurman, G. (2020). The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation. BMC Genomics, 21(1), 6. https://doi.org/10.1186/s12864-019-6413-7

[15] McNemar, Q. (1947). Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2), 153–157. https://doi.org/10.1007/BF02295996

[16] Bouteska, A., Abedin, M. Z., Hajek, P., & Yuan, K. (2024). Cryptocurrency price forecasting – A comparative analysis of ensemble learning and deep learning methods. International Review of Financial Analysis, 92, 103055. https://doi.org/10.1016/j.irfa.2023.103055




DOI: http://dx.doi.org/10.30970/eli.34.5

Refbacks

  • There are currently no refbacks.