Learning Predictive Ambiguity Sets for Decision-Focused Distributionally Robust Optimization

Battle Count: 73

Highly relevant to quantitative trading. The paper directly addresses portfolio optimization with 20 S&P 500 constituents, introduces adaptive robustness that responds to market volatility regimes, and demonstrates practical out-of-sample performance (26.28% annualized return, Sharpe 1.30). The regime analysis shows LPAS excels in drawdown and high-uncertainty periods, which are critical for risk-managed trading strategies. The decision-focused learning paradigm aligns with modern quant approaches that optimize for realized P&L rather than prediction accuracy alone.

Moderate-to-high complexity. Requires implementing: (1) a Transformer scenario generator with discrete output distribution, (2) a radius network with softplus/sigmoid parameterization, (3) a Wasserstein DRO decision layer with dual reformulation, (4) a multi-component training objective with staged training (pretrain, calibrate, fine-tune), (5) portfolio-specific constraints (long-only, simplex, transaction costs, covariance regularization). The DRO dual for the portfolio case has a closed-form (Eq. 20), simplifying the decision layer. However, the full pipeline with multiple loss terms, hyperparameter tuning, and regime-aware validation adds engineering overhead.

The paper provides detailed algorithm (Algorithm 1), explicit hyperparameters (ridge coefficient 1.0, risk penalty 8.0, transaction cost weight 0.0015, quantile level tau=0.9, radius floor 1e-4), model architecture (Transformer with 2 layers, 4 heads, hidden dim 48, 7 scenarios, lookback 63), and the full training objective. However, no code repository is provided, and the specific feature construction pipeline and data preprocessing steps are not fully detailed. The 20 S&P 500 tickers and date range are specified, enabling data reconstruction.