Decision-Induced Ranking Explains Prediction Inflation and Excessive Turnover in SPO-Based Portfolio Optimization

Battle Count: 76

Highly relevant to quantitative trading. The paper directly addresses practical challenges in ML-based portfolio optimization: prediction inflation, excessive turnover, and implementability of decision-focused learning strategies. The KKT-based ranking interpretation provides theoretical grounding for understanding SPO behavior. The stabilization mechanisms (clipping, rescaling, partial adjustment) are directly applicable to real trading systems. The empirical results on DOW and ETF universes with realistic transaction costs and monthly rebalancing are practical. The finding that increasing risk aversion alone cannot stabilize SPO portfolios is important for practitioners.

Moderate complexity. The linear predictor and MVO formulation are straightforward. However, implementing SPO+ surrogate training requires careful handling of the non-differentiable optimization layer, gradient computation through the optimizer, and proper hyperparameter tuning via Bayesian optimization. The stabilization interventions (clipping, min-max rescaling, partial portfolio adjustment) are simple post-processing steps. The KKT analysis is theoretical. The main complexity lies in correctly implementing the SPO+ training loop with the downstream MVO solver and ensuring numerical stability.

The paper provides detailed experimental configuration (Table 2) including datasets, hyperparameters, rolling window scheme, and model specifications. However, no code repository or exact implementation details for the SPO+ training loop are provided. The datasets (DOW, ETF_A, ETF_B) are publicly available. Five random seeds are used for robustness. The linear predictor and MVO formulation are standard, but the SPO+ surrogate training pipeline details are not fully specified.