End-to-End Parametric Portfolio Policies for Cross-Asset Futures Timing: When Do AI Models Beat Simple Rules?

Battle Count: 76

Directly addresses a core quantitative trading question: when does an AI-based allocation policy outperform simple rules (equal weight, risk parity, TSMOM) in a practical managed-futures/global-macro setting. Uses highly liquid CME futures (the actual instruments traded by systematic macro funds), realistic transaction costs, walk-forward validation, and practitioner-relevant benchmarks. The findings on turnover sensitivity, cost erosion, and the conditional nature of AI advantage are directly actionable for quantitative portfolio managers. The paper explicitly frames itself as a practitioner-oriented evaluation rather than an academic exercise.

The transformer architecture involves 4-layer encoder-decoder with multi-head self-attention, GRN, Time2Vec positional embeddings, causal masking, and a signed-softmax output layer. Training requires careful hyperparameter tuning (AdamW, OneCycleLR, GELU, pre-activation LayerNorm, gradient clipping, explicit initialization). The differentiable Sharpe ratio objective and walk-forward pipeline add complexity. Seed averaging (3 seeds) is needed for stability. However, the paper uses standard components and the LSTM baseline is ~30x faster to train. The main complexity lies in getting stable optimization and avoiding overfitting rather than novel architecture design.

The paper provides detailed architecture specifications (4-layer encoder-decoder transformer, GRN, Time2Vec, signed-softmax), training hyperparameters (AdamW, OneCycleLR, GELU, pre-activation LayerNorm, gradient clipping), walk-forward design (40% initial train, ~4-year test blocks, 90/10 train-validation split), and cost assumptions (2bp). However, no code repository is provided, Barchart data is commercial, and the paper notes seed sensitivity requiring 3-seed averaging. The training recipe was tuned primarily on the LSTM and verified on the transformer.