Research map / Image generation / Fast diffusion sampling

Retrieval-Augmented Diffusion Modeling for Stochastic Discount Factor Portfolios

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Diffusion theory and design space · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RADAR, a retrieval-augmented diffusion framework that uses retrieved historical news-event segments to set context-dependent noise statistics for denoising multimodal (price and news) market representations for SDF-based portfolio optimization.

Problem

Learning clean, regime-aware market representations for SDF-based portfolio optimization under non-stationary markets and noisy multimodal (price and news) inputs, where standard diffusion uses isotropic Gaussian noise.

Approach

Replaces isotropic Gaussian diffusion noise with noise whose mean and variance are the empirical statistics of top-K retrieved historical event segments (cosine similarity over a context bank), so the forward and reverse diffusion are grounded in state-dependent uncertainty; the denoised multimodal embeddings are trained jointly with an SDF pricing loss.

Key results

• RADAR achieves Sharpe 1.062, Sortino 1.352, Calmar 0.924 versus best baseline Equal Weight Sharpe 0.795 and StockNet Sortino 0.899 in the S&P 500 backtest.
• RADAR reports CAGR-style annualized return of 0.285 vs 0.156 for the best forecasting baseline (StockNet), but its max drawdown (0.308) is worse than MLP (0.093) and LinAttn (0.111).
• Adding Gaussian diffusion on top of embeddings drops Sharpe from 0.859 to 0.411, while RADAR's retrieval-grounded noise recovers to 1.062.

Surprising findings

• Adding standard Gaussian diffusion noise on top of learned embeddings substantially degrades performance (Sharpe 0.859 to 0.411), while retrieval-grounded noise improves it.
• Forecasting models with top-50 equal-weighted portfolios perform similarly to equal weighting regardless of forecasting quality, suggesting forecasting accuracy matters less in multi-asset portfolios.

Limitations

• Relies on the FNSPID news dataset, which is relatively sparse in coverage; retrieval quality depends on available historical news.
• Embedding-based portfolio weights are less interpretable than classical SDF approaches on raw characteristics.
• Gains are more limited in Energy and Real Estate sectors dominated by macro shocks not captured by firm-level data.

Experimental setup

Datasets: Yahoo Finance adjusted close prices FNSPID news Time-MMD

Benchmarks: S&P 500 stocks portfolio backtest (377 equities) Time-MMD

Metrics: Sharpe ratio Sortino ratio Calmar ratio cumulative return annualized return max drawdown volatility CAPM alpha Fama-French alpha Diebold-Mariano test

Baselines: Equal Weight S&P 500 HAN StockNet iTransformer NGAT RATD BSV DKKM MLP LinAttn SDF

Builds on: stochastic discount factor (SDF) portfolio model RATD denoising diffusion probabilistic models FNSPID

Closest papers by meaning

• In-Context Learning of Stochastic Differential Equations with Foundation Inference Models (NeurIPS 2025 · similarity 0.85)
• Towards Financial World Modeling (NeurIPS 2026 · similarity 0.85)
• Latent Variable Estimation in Bayesian Black-Litterman Models (ICML 2025 · similarity 0.84)
• DHMoE: Diffusion Generated Hierarchical Multi-Granular Expertise for Stock Prediction (AAAI 2025 · similarity 0.83)
• AlphaQCM: Alpha Discovery in Finance with Distributional Reinforcement Learning (ICML 2025 · similarity 0.83)
• Pareto-Optimality, Smoothness, and Stochasticity in Learning-Augmented One-Max-Search (ICML 2025 · similarity 0.83)