Smart Predict–then–Optimize Paradigm for Portfolio Optimization in Real Markets

Battle Count: 241

Highly relevant to quantitative trading. The paper directly addresses portfolio optimization with realistic trading frictions (transaction fees, turnover constraints), uses a practical rolling-window backtesting framework with monthly rebalancing, and demonstrates performance across multiple market regimes including crisis periods. The SPO paradigm provides a principled way to align ML predictions with actual trading decisions, which is a core challenge in quantitative finance.

Moderate complexity. The linear predictors are simple, but the SPO+ loss requires differentiating through an optimization layer (implemented via PyEPO). The robust SPO variant adds Monte Carlo sampling for worst-case scenarios. The rolling-window backtesting with Optuna hyperparameter tuning adds engineering complexity. The overall architecture is tractable but requires careful implementation of gradient propagation through the optimization layer.

The paper provides detailed model configurations (Table 2, 3), backtesting protocol, feature descriptions, and uses the open-source PyEPO library for SPO loss implementation. However, no GitHub repository is explicitly mentioned, and the exact ETF universe composition is not fully specified. Hyperparameter ranges are given but specific optimal values per window are not reported.