Diffolio: A Diffusion Model for Multivariate Probabilistic Financial Time-Series Forecasting and Portfolio Construction

Battle Count: 76

Highly relevant to quantitative trading. The paper directly addresses multivariate probabilistic forecasting for portfolio construction, with explicit evaluation through mean-variance tangency and growth-optimal portfolio strategies. Achieves Sharpe ratio of 0.7206 (40%+ higher than baselines and market). Provides calibrated uncertainty estimates critical for risk management. Demonstrates robustness across market regimes including financial crises. Inference speed (~1.2s for 100 samples per time step) supports practical daily/weekly rebalancing. The hierarchical architecture inspired by asset pricing theory (IPCA, factor models) provides economically meaningful forecasts.

Moderate-to-high complexity. Requires implementing: (1) DDPM/DDIM diffusion framework with noise scheduling, (2) two-stage hierarchical attention (cross-attention for asset-level, self-attention for market-level), (3) correlation-guided regularizer with Ledoit-Wolf shrinkage estimation, (4) conditional generation with covariate integration, (5) portfolio optimization (MVP and GOP). The architecture is well-described with equations and pseudocode. Training requires GPU (RTX A6000 used), ~3 hours for 100K steps. Inference is fast (~2.5 min for full test period). Key challenge is proper implementation of the hierarchical attention and correlation regularizer interaction.

The paper provides detailed hyperparameter configurations, training procedures, and evaluation metrics. Data sources (Kenneth French's data library, Amit Goyal's website) are publicly available. Algorithm pseudocode is provided. However, no GitHub repository is explicitly linked, and data availability is stated as 'will be made available on request.' The architecture details, loss functions, and experimental setup are thoroughly documented.