Factor-Based Diffusion Model Enhances Contextual Portfolio Optimization

The paper presents a factor-based conditional diffusion model for contextual portfolio optimization, using a Diffusion Transformer with token-wise conditioning to model the cross-sectional distribution of next-day returns. The authors train the generative model conditioned on high-dimensional, asset-specific factor vectors, then sample from the learned conditional distribution to perform daily mean-variance and mean-CVaR portfolio optimization under transaction costs and realistic constraints. Empirical tests on the Chinese A-share market report consistent outperformance across multiple risk-adjusted metrics, and the work includes a formal error propagation analysis linking model approximation error to optimization outcomes.

The core modeling contribution is a conditional diffusion architecture that associates each asset with its own factor token while retaining cross-asset interaction via a Diffusion Transformer backbone. Key technical components:

• •token-wise conditioning so each asset's next-day return is explicitly conditioned on an asset-specific factor vector
• •generative sampling to produce scenario ensembles used by mean-variance and mean-CVaR optimizers
• •incorporation of transaction cost models and practical portfolio constraints during optimization
• •theoretical bounds that quantify how conditional distribution approximation errors affect portfolio-level risk and return estimates

The authors report experimental comparisons to standard benchmarks on A-share data, with improvements in risk-adjusted returns reported across multiple evaluation windows.

Applying generative diffusion models to high-dimensional, cross-sectional financial prediction is a growing trend; this paper moves beyond marginal return forecasting to learn joint conditional distributions, which is crucial when downstream decisions require scenario-aware optimization. The token-wise conditioning pattern is practically important because financial factor sets are asset-specific, and the Diffusion Transformer choice scales better than naive joint-density estimators in higher dimensions. The theoretical error analysis matters for practitioners because it provides a bridge between model mis-specification and portfolio performance, enabling more disciplined model selection and risk budgeting when using generative forecasts.

Validate robustness out of sample, especially under regime shifts and transaction frictions. Next steps include exploring alternative conditioning sets, stress testing the learned conditional distribution, and applying the method to other markets or multi-asset universes.