MACRO-AWARE TIME SERIES FORECASTING VIA HIERARCHICAL MIXED-FREQUENCY ATTENTION MODELS

Battle Count: 87

Highly relevant to quantitative trading. The paper directly addresses systematic multi-asset trading with time-series momentum and carry strategies across 55 liquid futures. It provides concrete performance improvements (Sharpe ratio nearly doubling from 1.19 to 1.99 for momentum), transaction cost sensitivity analysis, and interpretable regime detection. The mixed-frequency attention mechanism is a novel contribution to the deep learning for finance literature, bridging macro-finance theory with practical trading implementation.

The architecture involves multiple components: LSTM encoders for macro and daily streams, hierarchical cross-attention with learnable temperature, self-attention blocks, PCA for macro dimensionality reduction, and a decoder LSTM. The mixed-frequency alignment (monthly attention projected to daily grid) adds complexity. However, the paper provides detailed architectural specifications and hyperparameters. Training requires GPU support with mixed precision. The 100-seed ensemble approach adds computational cost.

The paper provides detailed experimental configuration (Table 11), dataset descriptions (FRED-MD, Pinnacle), and architectural specifications. However, no GitHub repository is mentioned, and the Pinnacle database is proprietary. The FRED-MD data is publicly available. Walk-forward validation protocol and hyperparameters are documented. 100 seed trials with top 10% ensemble are specified.