CryptoGAT: Are Time Series Models Effective for Cryptocurrency Forecasting?

Battle Count: 215

Highly relevant to quantitative trading in cryptocurrency markets. The paper directly addresses price prediction for trading decisions, provides portfolio-level evaluation with Sharpe ratios, transaction cost analysis, and a concrete trading protocol. The finding that cross-asset graph structure outperforms temporal modeling has significant implications for crypto trading strategy design. The 3.128 annualized Sharpe ratio and cost sensitivity analysis make findings directly actionable for practitioners.

The proposed CryptoGAT model is described as 'simple yet effective' with only ~47K-53K parameters. The graph construction via Pearson correlation is straightforward (Algorithm 1). The GAT attention mechanism is standard. Training uses PyTorch with standard hyperparameters. The main complexity lies in data preprocessing (filtering, normalization, time alignment) rather than the model itself. Inference time is competitive (0.627-0.792ms).

Source code is publicly available on GitHub (https://github.com/FanBroWell/CryptoGAT). Detailed hyperparameter settings are provided in Table VI. Experiments use a well-known exchange (Binance) with clear data filtering criteria. However, the dataset spans 2023-2026 which may limit external reproducibility. All experiments repeated 3 times with average reported. Trading protocol (Algorithm 2) and graph construction (Algorithm 1) are fully specified.