Forecasting Economically Significant Bitcoin Moves: A Multi-Scale TCN with Profit-Optimized Thresholds

Battle Count: 159

Highly relevant to quantitative trading. The paper directly addresses a practical trading question (will Bitcoin exceed 5% in 7 days?), uses a profit-optimized decision threshold aligned with trading economics, incorporates transaction costs, employs walk-forward validation mimicking live trading, and integrates multiple data modalities (on-chain, market, sentiment) commonly used in crypto trading. The pairwise ranking loss and AUC-based evaluation are well-suited for imbalanced financial classification. However, the modest AUC (0.63) and lack of position sizing/execution details limit immediate deployability.

Moderate to high complexity. The architecture involves multiple components: InceptionTCN blocks with parallel dilated convolutions, bottleneck-fusion layers, CNN channel attention, adaptive average pooling, and a custom pairwise ranking loss. Training requires GPU resources, Optuna for hyperparameter tuning, and careful implementation of the walk-forward validation scheme. The multi-modal data pipeline (on-chain, market, sentiment) adds preprocessing complexity. However, all components are based on well-known building blocks (TCN, attention, pooling) and implemented in PyTorch.

The paper provides detailed architecture descriptions, hyperparameters (Table 6), data preprocessing steps, and evaluation protocols. However, no GitHub repository is mentioned, and the exact data sources (on-chain metrics from specific providers) are not fully specified. The walk-forward validation scheme is described but code is not referenced. Feature list is provided (Table 4) but exact data pipeline code is absent.