DeePM: Regime-Robust Deep Learning for Systematic Macro Portfolio Management

Battle Count: 81

This paper is directly and fundamentally about systematic macro portfolio management. It addresses core quantitative trading challenges: asynchronous data, transaction costs, regime shifts, and risk-adjusted return optimization. The empirical results (Net Sharpe 0.93 vs 0.45 for TSMOM) are evaluated on 50 diversified futures with realistic costs over 2010-2025. The architecture is designed for practical deployment with strict realizability constraints. The paper bridges academic deep learning and institutional CTA practice.

The architecture integrates multiple complex components: Vectorized Variable Selection Network with FiLM conditioning, LSTM with asset-specific initialization, Temporal Multi-Head Attention with SwiGLU adapters, Cross-Sectional Attention with Directed Delay and ReZero gating, Graph Attention Network with structural edge biases, and a non-separable Sharpe-based loss requiring exact two-pass gradient accumulation. The training protocol involves 50-100 seeds, walk-forward validation, burn-in periods, and ensemble selection. The transaction cost model requires asset-specific microstructure parameters. Implementation requires careful handling of permutation equivariance, existence masking, and filtration compliance.

The paper provides detailed hyperparameter search spaces (Table 5), architecture specifications, data processing pipelines, and optimization protocols. However, the data source (Pinnacle Data Corp CLC Database) is proprietary and not publicly available. No code repository is mentioned. The walk-forward validation scheme, transaction cost model, and macro graph topology are fully specified. Training on a single RTX 3090 GPU is noted. The 50-seed ensemble protocol and exact gradient accumulation method are described in detail.