Paper: arXiv 2508.01880
Authors: Duo Zhang, Jiayu Li, Junyi Mo, Elynn Chen
Abstract
Accurate volatility forecasts are vital in modern finance for risk management, portfolio allocation, and strategic decision-making. However, existing methods face key limitations. Fully multivariate models, while comprehensive, are computationally infeasible for realistic portfolios. Factor models, though efficient, primarily use static factor loadings, failing to capture evolving volatility co-movements when they are most critical. To address these limitations, we propose a novel, model-agnostic Factor-Augmented Volatility Forecast framework. Our approach employs a time-varying factor model to extract a compact set of dynamic, cross-sectional factors from realized volatilities with minimal computational cost. These factors are then integrated into both statistical and AI-based forecasting models, enabling a unified system that jointly models asset-specific dynamics and evolving market-wide co-movements. Our framework demonstrates strong performance across two prominent asset classes-large-cap U.S. technology equities and major cryptocurrencies-over both short-term (1-day) and medium-term (7-day) horizons. Using a suite of linear and non-linear AI-driven models, we consistently observe substantial improvements in predictive accuracy and economic value. Notably, a practical pairs-trading strategy built on our forecasts delivers superior risk-adjusted returns and profitability, particularly under adverse market conditions.
Complexity vs Empirical Score
- Math Complexity: 7.0/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced statistical and machine learning techniques (e.g., time-varying factor models, LSTM) requiring dense mathematical formulation, while also demonstrating strong empirical validation through backtested trading strategies and detailed performance metrics on real-world datasets.
Research Flowchart
flowchart TD A["Research Goal<br>Improve Volatility Forecasting<br>for Risk Management"] --> B["Data Inputs<br>Large-cap Tech Equities &<br>Major Cryptocurrencies"] B --> C["Methodology<br>Time-Varying Factor Model"] C --> D["Computational Process<br>Extract Dynamic Market Factors"] D --> E["Integration<br>Augment AI & Statistical Models"] E --> F["Key Findings<br>Sustained Predictive Accuracy<br>& Economic Value"] E --> G["Outcome<br>Superior Pairs-Trading Strategy<br>High Risk-Adjusted Returns"]
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 a novel, model-agnostic Factor-Augmented Volatility Forecast framework.
Source sentence
we propose a novel, model-agnostic Factor-Augmented Volatility Forecast framework.
Assumes 4
- Daily position sizes are scaled to achieve a 25% annualized volatility target. execution
Source sentence
Our strategy employs a volatility-targeting approach, where daily position sizes are scaled to achieve a 25% annualized volatility target.
- A transaction cost of 5 basis points per leg is applied in the backtest. execution
Source sentence
The backtest is initialized with an equity of $50,000 and incorporates a 5 basis point transaction cost per leg.
- Trading signals are generated from a rolling z‑score of the spread with thresholds ±1.5. execution
Source sentence
Trading signals are generated from a rolling z-score of the spread, where 𝑍𝑡=𝑆𝑡−𝜇𝑆𝑡 / 𝜎𝑆𝑡 . We initiate a long position when 𝑍𝑡< -1.5 and a short position when 𝑍𝑡> +1.5, closing the position when the z-score reverts to zero.
- Statistical models use a 70‑day rolling window; LSTM uses a 30‑day window. model
Source sentence
The rolling window length (𝑊) is set to 70 days for the statistical models to balance responsiveness to regime changes with statistical stability, while a shorter window of 30 days is used for the LSTM to align with its input structure.
Data
cryptocurrencies · crypto · daily
- Equity high‑frequency dataset market-prices commercial
Source sentence
The equity dataset focuses on five prominent firms within the S&P 500 Information Technology Sector, selected for their large market capitalization and historical sectoral comovements: Microsoft (MSFT), Advanced Micro Devices (AMD), Intel (INTC), Oracle (ORCL), and Cisco (CSCO). For these firms, we obtained high-freque
- Cryptocurrency high‑frequency dataset crypto-onchain commercial
Source sentence
The cryptocurrency dataset, sourced from Kaiko, covers January 2018 to April 2021. This dataset comprises detailed, high-frequency trading data for five leading cryptocurrencies: Bitcoin (BTC), Ethereum (ETH), Ripple (XRP), Cardano (ADA), and Litecoin (LTC).
Kaiko
Costs
Modeled
5 basis point transaction cost per leg
Source sentence
The backtest is initialized with an equity of $50,000 and incorporates a 5 basis point transaction cost per leg.
Finds 3
Sharpe 1.49 · annual return 7.3%
- Factor augmentation reduces forecast errors by up to 12.8% for equities and 22.8% for cryptocurrencies, lowering MSE and QLIKE. predictive-accuracy
Source sentence
eases by up to 12.8% for equities and 22.8% for cryptocurrencies, with corresponding reductions in Mean Squared Error (MSE) and Quasi-likelihood loss (QLIKE).
- AR model’s annualized return improves by 23.6% after factor augmentation (5.35% → 6.61%) and Sharpe ratio rises from 1.437 to 1.443. performance
Source sentence
... specification boosts its annualized return by a remarkable 23.6%.
- Factor‑augmented LSTM flips a -5.5% annualized loss into a +7.3% gain and improves Sharpe from -0.467 to 0.787 in a challenging market period. performance
Source sentence
In a stunning reversal, our factor-augmented LSTM transformed this failing strategy into a resounding success, flipping the annualized loss into a +7.3% gain. This phenomenal turnaround propelled its Sharpe ratio from a negative -0.467 to a positive 0.787.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 30 related papers). Coverage is partial, so this is neither replication nor refutation.
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 (111)
- Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation (1982)
- Generalized autoregressive conditional heteroskedasticity (1986)
- A simple approximate long-memory model of realized volatility (2009)
- Long short-term memory (1997)
- Deep learning with long short-term memory networks for financial market predictions (2018)
- Multi-transformer: A new neural network-based architecture for forecasting S&P volatility (2021)
- Volatility forecasting and assessing risk of financial markets using multi-transformer neural network based architecture (2024)
- Transformer-based attention network for stock movement prediction (2022)
and 103 more
- Alternative models for conditional stock volatility (1990)
- A conditionally heteroskedastic time series model for speculative prices and rates of return (1987)
- An evaluation of volatility forecasting techniques (1996)
- Volatility forecasting for risk management (2003)
- Conditional heteroskedasticity in asset returns: A new approach (1991)
- RiskMetrics\textsuperscript \textregistered —Technical Document (1996) — link
- Volatility forecasting via MIDAS, HAR and their combination: An empirical comparative study for IBOVESPA (2014)
- Two are better than one: volatility forecasting using multiplicative component GARCH-MIDAS models (2020)
- Long memory processes and fractional integration in econometrics (1996)
- Persistence in variance, structural change, and the GARCH model (1990)
- A multiple regime smooth transition heterogeneous autoregressive model for long memory and asymmetries (2008)
- Structural breaks and volatility forecasting in the copper futures market (2018)
- GARCH-Informed Neural Networks for Volatility Prediction in Financial Markets — → 2410.00288
- An artificial neural network-GARCH model for international stock return volatility (1997)
- Volatility forecast using hybrid neural network models (2014)
- Realized volatility forecasting with neural networks (2020)
- Experiments in predicting the German stock index DAX with density estimating neural networks (1996)
- Application of neural network for forecasting of exchange rates and forex trading (2012)
- A LSTM-based method for stock returns prediction: A case study of China stock market (2015)
- Novel volatility forecasting using deep learning–long short term memory recurrent neural networks (2019)
- Stock market prediction using LSTM recurrent neural network (2020)
- Forecasting the volatility of stock price index: A hybrid model integrating LSTM with multiple GARCH-type models (2018)
- Risk everywhere: Modeling and managing volatility (2018)
- Modeling and forecasting large realized covariance matrices and portfolio choice (2017)
- A Neural Network with Shared Dynamics for Multi‑Step Prediction of Value‑at‑Risk and Volatility (2022) — link
- Volatility forecasting with machine learning and intraday commonality (2024)
- Neural Network–Based Financial Volatility Forecasting: A Systematic Review (2023) — link
- Predicting the Price of Bitcoin Using Machine Learning (2018) — link
- Predicting Cryptocurrency Market Volatility with Deep Learning Chaos Networks (2019) — link
- Bitcoin Is Not the New Gold: A Comparison of Volatility, Correlation and Portfolio Performance (2018) — link
- An Advanced CNN–LSTM Model for Cryptocurrency Forecasting (2021) — link
- Technical analysis and sentiment embeddings for market trend prediction (2019)
- A hybrid volatility forecasting framework integrating GARCH, artificial neural network, technical analysis and principal components analysis (2018)
- Forecasting realized volatility: The role of implied volatility, leverage effect, overnight returns, and volatility of realized volatility (2021)
- The Distribution of Realized Stock Return Volatility (2001) — link
- MIDAS Regressions: Further Results and New Directions (2007) — link
- Comparing predictive accuracy (2002)
- Volatility forecast comparison using imperfect volatility proxies (2011)
- Attention is all you need (2017)
- Correlations in price changes and volatility across international stock markets (1990)
- The common factor in idiosyncratic volatility: Quantitative asset pricing implications (2016)
- Volatility spillover in crypto-currency markets: Some evidences from GARCH and wavelet analysis (2019)
- Multivariate simultaneous generalized ARCH (1995)
- Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models (2002)
- The conditional autoregressive Wishart model for multivariate stock market volatility (2012)
- A vector heterogeneous autoregressive index model for realized volatility measures (2017)
- Multivariate GARCH models: a survey (2006)
- Large dynamic covariance matrices (2019)
- GO-GARCH: a multivariate generalized orthogonal GARCH model (2002)
- A primer on the orthogonal GARCH model (2000)
- Dynamic conditional beta (2016)
- Deep learning in characteristics-sorted factor models (2024)
- Time-Varying Matrix Factor Models (2024) — link
- Eigenvalue ratio test for the number of factors (2013)
- Determining the number of factors from empirical distribution of eigenvalues (2010)
- The spline GARCH model for unconditional volatility and its global macroeconomic causes (2005)
- Empirical asset pricing via machine learning (2020)
- A factor approach to realized volatility forecasting in the presence of finite jumps and cross-sectional correlation in pricing errors (2013)
- Common risk factors in the returns on stocks and bonds (1993)
- On persistence in mutual fund performance (1997)
- Consistent estimation of time-varying loadings in high-dimensional factor models (2019)
- Multiplicative factor model for volatility (2025)
- Dynamic factor models with time-varying parameters: measuring changes in international business cycles (2008)
- Global market integration: An alternative measure and its application (2009)
- Principal component analysis: a review and recent developments (2016)
- Statistical inference for high-dimensional matrix-variate factor models (2023)
- On projection robust optimal transport: Sample complexity and model misspecification (2021)
- Identification and estimation of threshold matrix-variate factor models (2022)
- Community network auto-regression for high-dimensional time series (2023)
- Constrained factor models for high-dimensional matrix-variate time series (2019)
- Modeling dynamic transport network with matrix factor models: with an application to international trade flow (2022)
- Modeling Multivariate Spatial-Temporal Data with Latent Low-Dimensional Dynamics (2020) — link
- Transfer Q-learning (2022) — link
- Reinforcement Learning in Latent Heterogeneous Environments (2024)
- Tensor-view Topological Graph Neural Network (2024)
- MEV Makes Everyone Happy under Greedy Sequencing Rule (2023)
- Semi-parametric tensor factor analysis by iteratively projected singular value decomposition (2024)
- Time-varying Matrix Factor Models (2024)
- Computation of optimal mev in decentralized exchanges (2024)
- Data-Driven Knowledge Transfer in Batch $Q^ *$ Learning (2024) — link
- Dynamic Contextual Pricing with Doubly Non-Parametric Random Utility Models (2024) — link
- Factor Augmented Matrix Regression (2024) — link
- High-Dimensional Tensor Classification with CP Low-Rank Discriminant Structure (2024) — link
- Advancing Information Integration through Empirical Likelihood: Selective Reviews and a New Idea (2024) — link
- Conditional Prediction ROC Bands for Graph Classification (2024)
- TEAFormers: TEnsor-Augmented Transformers for Multi-Dimensional Time Series Forecasting (2024)
- Exploring causal effects of hormone-and radio-treatments in an observational study of breast cancer using copula-based semi-competing risks models (2024) — link
- Deep Transfer $ Q $-Learning for Offline Non-Stationary Reinforcement Learning (2025) — link
- Statistical Inference for Low-Rank Tensor Models (2025) — link
- Stochastic Linear Bandits with Latent Heterogeneity (2025) — link
- Bridging Domain Adaptation and Graph Neural Networks: A Tensor-Based Framework for Effective Label Propagation (2025)
- SEMIPARAMETRIC TENSOR FACTOR ANALYSIS BY ITERATIVELY PROJECTED SVD BY ELYNN Y. CHEN, DONG XIA, CHENCHENG CAI, AND JIANQING FAN (2020) — link
- Distributed Tensor Principal Component Analysis with Data Heterogeneity (2025)
- Transfer Faster, Price Smarter: Minimax Dynamic Pricing under Cross-Market Preference Shift (2025)
- Transfer Q-Learning with Composite MDP Structures (2025)
- Maximal Extractable Value in Batch Auctions (2025)
- PostDoctoral Research Fellowship (2018)
- Low-Rank Principal Eigenmatrix Analysis (2019) — link
- High-Dimensional Structured Bandits: Minimax Regret, Estimation and Inference (2020)
- Collaborative Research: High-Dimensional Tensor Learning under Labeled-Data Scarcity (2024)
- Time-Varying Factor-Augmented Models for Volatility Forecasting (2025)
- ACT-Tensor: Tensor Completion Framework for Financial Dataset Imputation (2025)
- Tensor-fused multi-view graph contrastive learning (2025)
1 of 111 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.