Learning Predictive Ambiguity Sets for Decision-Focused Distributionally Robust Optimization

Technical Summary: Learning Predictive Ambiguity Sets for Decision-Focused Distributionally Robust Optimization

1. Problem Statement

Standard "predict-then-optimize" systems compress uncertainty into a single point forecast (e.g., a mean return) and solve a downstream optimization problem as if that forecast were ground truth. This approach is fragile; small prediction errors can lead to large decision losses, particularly when the optimizer amplifies mistakes along active constraints or high-sensitivity directions.

Distributionally Robust Optimization (DRO) addresses this by optimizing against the worst-case distribution within an ambiguity set. However, traditional Wasserstein DRO relies on hand-designed ambiguity sets where the center is a fixed historical empirical distribution and the radius is a global scalar tuned via validation. This static approach is often mismatched to contextual environments: a radius safe during volatile periods may be overly conservative during stable periods, while a radius tuned for average performance may fail under regime shifts. The core problem is how to make the ambiguity set itself contextual and learnable, adapting to the specific state of the system to balance robustness and performance.

2. Methodology: Learned Predictive Ambiguity Sets (LPAS)

The paper proposes Learned Predictive Ambiguity Sets (LPAS), a framework that bridges probabilistic deep forecasting and Wasserstein DRO. Instead of a fixed center and radius, LPAS employs a deep contextual model that outputs three components for a given context $z_t$:

1. A Finite Nominal Scenario Distribution: $\hat{P}_\theta(\cdot | z_t) = \sum_{i=1}^N p_{\theta,i}(z_t) \delta_{\hat{\xi}_{\theta,i}(z_t)}$, representing a discrete set of predicted scenarios.
2. A State-Dependent Radius: $\rho_\phi(z_t)$, a non-negative scalar that defines the size of the Wasserstein ball.
3. (Optional) An Anisotropic Ground Metric: $c_\psi(\xi, \xi'; z_t)$, though the experiments utilize a fixed Euclidean metric.

These outputs define a contextual ambiguity set $\mathcal{P}_{\theta,\phi}(z_t) = \{ Q : W_c(Q, \hat{P}_\theta(\cdot | z_t)) \leq \rho_\phi(z_t) \}$. The downstream decision $x_t$ is computed by minimizing the worst-case expected loss over this set.

Training Objective and Algorithm
The model is trained end-to-end using a composite objective function that balances four terms:

• Decision Loss ($L_{dec}$): The realized loss of the downstream robust decision.
• Prediction Loss ($L_{pred}$): Trains the nominal scenario distribution (e.g., via negative log-likelihood or scenario reconstruction).
• Calibration Loss ($L_{cal}$): Uses pinball loss to ensure the radius approximates a conditional high quantile of the forecast errors, preventing the radius from collapsing.
• Size and Stability Regularization ($L_{size}, L_{stab}$): Prevents radius inflation and ensures smooth transitions between time steps.

The training follows a staged algorithm:

1. Pretrain the scenario generator ($\hat{P}_\theta$) using prediction loss.
2. Train the radius network ($\rho_\phi$) using calibration and size regularization on a calibration split.
3. Fine-tune the entire system (scenarios, radius, and decision layer) using the full decision-focused objective.

For the specific case of portfolio optimization with affine loss and Euclidean costs, the paper derives a tractable finite-dimensional dual form. The robust portfolio problem becomes a maximization of expected return minus a penalty term proportional to the learned radius and the portfolio norm: $\max_w \hat{\mu}^\top w - \rho(z_t)\|w\|_2 - \text{regularization}$.

3. Key Contributions

The paper makes four primary contributions:

1. LPAS Framework: Introduces a contextual bridge between deep probabilistic forecasting and Wasserstein DRO, where the ambiguity set (center and radius) is learned rather than fixed.
2. Radius-Learning Objective: Develops a multi-component loss function that integrates prediction accuracy, statistical calibration (quantile-style), size regularization, and realized decision quality.
3. Tractable Implementation: Provides a finite dual form for the DRO layer and a staged training algorithm that enables differentiable robust decision-making.
4. Empirical Validation: Demonstrates via a portfolio optimization study that adaptive ambiguity radii can outperform non-robust and historical-DRO baselines, while matching much of the performance of a strong fixed-radius DRO model with a significantly smaller average radius.

4. Experimental Results

The method was evaluated on a daily portfolio optimization task using 20 S&P 500 constituents from January 2018 to June 2026. The LPAS model was compared against Equal Weight (EW), Predict-then-Optimize (P2O), Historical Wasserstein DRO (Hist-WDRO), and a Deep Predictive Fixed-Radius DRO (Fixed-DRO).

Key Findings:

• Performance: LPAS substantially outperformed EW, P2O, and Hist-WDRO.
• Comparison to Fixed-DRO: LPAS achieved an annualized return of 26.28% and a Sharpe ratio of 1.30, slightly lower than the Fixed-DRO baseline (27.48% return, 1.34 Sharpe).
• Efficiency and Risk: Despite the slight drop in average metrics, LPAS used a substantially smaller average radius (24.35 vs. 35.44) and achieved better worst-month loss and CVaR95 metrics.
• Adaptivity: The learned radius successfully varied with market conditions, increasing during high-volatility regimes and decreasing during stable periods.
• Ablation: Removing the calibration loss caused the radius to collapse, leading to negative returns and high drawdowns, confirming that calibration is essential for the method's stability.

5. Significance and Claims

The paper claims that the proposed approach demonstrates that learned ambiguity radii can recover most of the performance of strong fixed-radius DRO while reducing unnecessary conservatism.

The central hypothesis supported by the results is that in robust decision-making, the optimal object to learn is not merely a prediction, but also how much the optimizer should distrust that prediction based on the current state. By making the ambiguity set contextual, LPAS achieves "regime adaptivity," providing stronger protection during drawdowns and high-uncertainty periods without sacrificing performance in calmer markets. The authors note that while LPAS is competitive with fixed-radius DRO, it offers a more efficient use of the robustness budget, avoiding the global over-conservatism inherent in fixed-radius approaches.

Limitations Acknowledged:
The authors modestly note that the experiments are limited to a single rolling split on a 20-asset universe, the learned anisotropic metric was not evaluated, and empirical coverage is slightly below the nominal target after decision-aware tuning (suggesting a need for conformal post-calibration in strict risk-control applications).