Paper: arXiv 2511.07014
Authors: So-Yoon Cho, Jin-Young Kim, Kayoung Ban, Hyeng Keun Koo, Hyun-Gyoon Kim
Abstract
Probabilistic forecasting is crucial in multivariate financial time-series for constructing efficient portfolios that account for complex cross-sectional dependencies. In this paper, we propose Diffolio, a diffusion model designed for multivariate financial time-series forecasting and portfolio construction. Diffolio employs a denoising network with a hierarchical attention architecture, comprising both asset-level and market-level layers. Furthermore, to better reflect cross-sectional correlations, we introduce a correlation-guided regularizer informed by a stable estimate of the target correlation matrix. This structure effectively extracts salient features not only from historical returns but also from asset-specific and systematic covariates, significantly enhancing the performance of forecasts and portfolios. Experimental results on the daily excess returns of 12 industry portfolios show that Diffolio outperforms various probabilistic forecasting baselines in multivariate forecasting accuracy and portfolio performance. Moreover, in portfolio experiments, portfolios constructed from Diffolio’s forecasts show consistently robust performance, thereby outperforming those from benchmarks by achieving higher Sharpe ratios for the mean-variance tangency portfolio and higher certainty equivalents for the growth-optimal portfolio. These results demonstrate the superiority of our proposed Diffolio in terms of not only statistical accuracy but also economic significance.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 6.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced mathematics including diffusion models, hierarchical attention networks, and a correlation-guided regularizer with derivations. Empirical evaluation uses daily excess returns on 12 industry portfolios, reporting Sharpe ratios and certainty equivalents, but lacks code or detailed implementation details.
Research Flowchart
flowchart TD A["Research Goal<br>Multivariate Probabilistic Financial<br>Time-Series Forecasting & Portfolio Construction"] --> B["Data & Inputs<br>Historical Daily Excess Returns<br>Asset & Market Covariates"] B --> C["Methodology: Diffolio Model<br>Diffusion Process + Denoising<br>Hierarchical Attention: Asset & Market Levels"] C --> D["Key Component<br>Correlation-Guided Regularizer<br>Using Stable Target Correlation Matrix"] D --> E["Computational Processes<br>Forward/Reverse Diffusion<br>Network Training & Inference"] E --> F["Key Outcomes<br>1. Superior Multivariate Forecasting Accuracy<br>2. Efficient Portfolio Construction"] F --> G["Final Results<br>Higher Sharpe Ratios & Certainty Equivalents<br>vs. Benchmarks"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes Diffolio, a diffusion-based model with hierarchical attention for multivariate financial time-series forecasting and portfolio construction.
Source sentence
In this paper, we propose Diffolio, a diffusion model designed for multivariate financial time-series forecasting and portfolio construction.
Assumes 6
- Portfolio optimization is performed without transaction cost constraints in the primary experiments. execution
Source sentence
Since the portfolio optimization process is conducted without transaction cost constraints, the models focus exclusively on the predicted joint distribution to determine optimal weights at each time step.
- A stable target correlation matrix can be estimated via Ledoit‑Wolf shrinkage and used to guide attention. model
Source sentence
we introduce a correlation-guided regularizer that employs a stable target correlation matrix estimated via the Ledoit‑Wolf shrinkage method.
- Daily excess returns of industry portfolios are forecast one day ahead. data
Source sentence
we forecast one-day-ahead daily excess returns of 12 industry portfolios
- Asset‑specific and systematic covariates are available and informative for forecasting. data
Source sentence
For the model’s covariates, we employ asset characteristics as the asset-specific variables and macroeconomic variables as the systematic variables
Show 2 more
- The market can be rebalanced weekly with a flat transaction cost of 10 basis points. execution
Source sentence
we explicitly integrate a transaction cost of 10 basis points (bps) as a penalty term directly into the portfolio optimization objective function and adjust the rebalancing frequency to a five-day interval (weekly)
- The deterministic DDIM sampling procedure provides efficient generation of return samples. model
Source sentence
For inference, we employ the deterministic DDIM sampling procedure described in Section 3.2, which allows for efficient generation of return samples in a reduced number of steps.
Data
U.S. industry portfolios · equities · daily
- 12 industry portfolios market-prices public
Source sentence
We forecast one-day-ahead daily excess returns of 12 industry portfolios
Costs
Modeled
10 basis points per trade, weekly (five‑day) rebalancing
Source sentence
we explicitly integrate a transaction cost of 10 basis points (bps) as a penalty term directly into the portfolio optimization objective function and adjust the rebalancing frequency to a five-day interval (weekly)
Finds 4
Sharpe 0.72 (benchmark 0.51) · annual return 16.8% · max drawdown 55.8%
- Diffolio outperforms baselines in multivariate forecasting accuracy and portfolio performance, achieving higher Sharpe ratios and certainty equivalents. performance
Source sentence
Diffolio outperforms various probabilistic forecasting baselines in both statistical multivariate forecasting accuracy and portfolio performance
- Diffolio attains the best Energy Score (ES) and second‑best CRPS, with the lowest CRPS standard deviation across assets. predictive-accuracy
Source sentence
the lowest standard deviation in CRPS across the 12 assets.
- In a frictionless setting, most models exhibit high turnover, but Diffolio remains responsive to daily market signals. other
Source sentence
... high turnover is incidental to the primary objectives of maximizing MVP‑SR or GOP‑CE; it serves as a measure of the model’s responsiveness.
- When realistic transaction costs (10 bps) and weekly rebalancing are applied, Diffolio still delivers superior portfolio metrics compared to benchmarks. performance
Source sentence
Following the portfolio results in the frictionless environment, we assess the practical viability
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 34 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (59)
- DeepAR: Probabilistic forecasting with autoregressive recurrent networks (2020)
- Capital market equilibrium with transaction costs (1986)
- Liquidity premia and transaction costs (2007)
- Time-series generative adversarial networks (2019)
- AlphaPortfolio: Direct construction through deep reinforcement learning and interpretable AI (2021)
- Timevae: A variational auto-encoder for multivariate time series generation (2021) — link
- Sig-Wasserstein GANs for conditional time series generation (2024)
- Diffwave: A versatile diffusion model for audio synthesis (2020) — link
and 51 more
- Csdi: Conditional score-based diffusion models for probabilistic time series imputation (2021)
- Diffusion-based time series imputation and forecasting with structured state space models (2022) — link
- Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting (2021)
- Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting (2023)
- Generative learning for financial time series with irregular and scale-invariant patterns (2024)
- Diffusion-ts: Interpretable diffusion for general time series generation (2024) — link
- Mg-tsd: Multi-granularity time series diffusion models with guided learning process (2024) — link
- Modeling financial time-series with generative adversarial networks (2019)
- Quant GANs: deep generation of financial time series (2020)
- Fin-gan: Forecasting and classifying financial time series via generative adversarial networks (2024)
- Factorvae: A probabilistic dynamic factor model based on variational autoencoder for predicting cross-sectional stock returns (2022)
- Generation of synthetic financial time series by diffusion models — → 2410.18897
- Wavenet: A generative model for raw audio (2016) — link
- A comprehensive look at the empirical performance of equity premium prediction (2008)
- Empirical asset pricing via machine learning (2020)
- Denoising diffusion probabilistic models (2020)
- Denoising diffusion implicit models (2020) — link
- Deep unsupervised learning using nonequilibrium thermodynamics (2015)
- Honey, I shrunk the sample covariance matrix (2003)
- Improved estimation of the covariance matrix of stock returns with an application to portfolio selection (2003)
- Hierarchical attention networks for document classification (2016)
- Characteristics are covariances: A unified model of risk and return (2019)
- Autoencoder asset pricing models (2021)
- Empirical properties of asset returns: stylized facts and statistical issues (2001)
- Portfolio Selection (1952)
- Probabilistic forecasting (2014)
- The dividend-price ratio and expectations of future dividends and discount factors (1988)
- Strictly proper scoring rules, prediction, and estimation (2007)
- Generative adversarial nets (2014)
- Auto-encoding variational bayes (2013) — link
- Flow matching for generative modeling (2023)
- Attention is all you need (2017)
- Layer normalization (2016) — link
- Perception encoder: The best visual embeddings are not at the output of the network (2025) — link
- Industry costs of equity (1997)
- Tree-based conditional portfolio sorts: The relation between past and future stock returns (2016)
- Deep learning in asset pricing (2024)
- Predicting residential energy consumption by explainable deep learning with long-term and short-term latent variables (2023)
- Explainable prediction of electric energy demand using a deep autoencoder with interpretable latent space (2021)
- Capital asset prices: A theory of market equilibrium under conditions of risk (1964)
- The arbitrage theory of capital asset pricing (1976)
- A financial time series denoiser based on diffusion models (2024)
- CoFinDiff: Controllable Financial Diffusion Model for Time Series Generation — → 2503.04164
- Estimation with quadratic loss (1961)
- Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models (2002)
- Betting against beta (2014)
- MarketGANs: Multivariate financial time-series data augmentation using generative adversarial networks — → 2601.17773
- Mgsfformer: A multi-granularity spatiotemporal fusion transformer for air quality prediction (2025)
- Ginar+: A robust end-to-end framework for multivariate time series forecasting with missing values (2025)
- Charging strategies optimization for lithium-ion battery: Heterogeneous ensemble surrogate model-assisted advanced multi-objective optimization algorithm (2025)
- MetaGNSDformer: Meta-learning enhanced gated non-stationary informer with frequency-aware attention for point-interval remaining useful life prediction of lithium-ion batteries (2026)
3 of 59 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.