Factor-Based Conditional Diffusion Model for Contextual Portfolio Optimization

Battle Count: 82

This paper is directly relevant to quantitative trading as it addresses contextual portfolio optimization - a core problem in quantitative finance. It proposes a novel generative approach to learning conditional return distributions for daily portfolio rebalancing, incorporates realistic transaction costs, and demonstrates superior performance over standard benchmarks (empirical estimators, shrinkage estimators, DCC-GARCH, equal-weight) in both mean-variance and mean-CVaR frameworks. The work bridges generative AI and practical portfolio construction with theoretical guarantees.

High complexity due to: (1) Modified DiT architecture with token-wise conditioning requiring custom implementation; (2) DDPM training with noise scheduling and reverse sampling; (3) Daily portfolio optimization solving quadratic programs (mean-variance) or linear programs (mean-CVaR) with transaction cost constraints; (4) Data preprocessing including cross-sectional standardization, winsorization, and missing value imputation; (5) Handling of suspended stocks and price limits; (6) Training on GPU with grid search hyperparameter optimization. However, the paper reports only ~20 minutes training time, suggesting the architecture is not excessively large.

The paper provides detailed experimental setup including data sources (Wind Database, Datayes), hyperparameter optimization via grid search, training time (~20 minutes on NVIDIA A30 GPU), and specific factor descriptions (208 price-volume factors). However, no GitHub repository or code is mentioned. The DiT architecture modifications are described in detail. Data sources are commercial and may not be freely accessible.