Deep Learning Study Compares Chart Types for Crypto Regime Prediction
A detailed study has been carried out using deep learning techniques to forecast cryptocurrency trends by analyzing candlestick charts. Researchers looked at three ways to encode images—raw candlestick charts, Gramian Angular Fields, and multi-channel GAF—along with five chart component setups and four types of neural networks, including CNN, ResNet18, EfficientNet-B0, and Vision Transformer, while also considering ImageNet transfer learning. Over eight controlled experiments involving Bitcoin, Ethereum, and S&P 500 data from 2018 to 2024, they identified the best configurations. Impressively, a simple 4-layer CNN with raw candlestick charts achieved an AUC-ROC of 0.892, outperforming more complex models.
Key facts
- Study compares visual representations for cryptocurrency regime prediction
- Evaluates three image encoding methods: raw candlestick charts, Gramian Angular Fields, multi-channel GAF
- Tests five chart component configurations and four neural network architectures
- Architectures include CNN, ResNet18, EfficientNet-B0, and Vision Transformer
- Examines impact of ImageNet transfer learning
- Eight controlled experiments on Bitcoin, Ethereum, and S&P 500 data from 2018-2024
- Simple 4-layer CNN on raw candlestick charts achieves 0.892 AUC-ROC
- Outperforms more complex models in visual regime classification
Entities
Institutions
- arXiv