💻 computer science

Forecasting Global Volatility Across Asynchronous Markets: Incremental Accuracy from Constrained Cross-Market Attention

The paper introduces PGA-Trans-HAR, a novel forecasting framework that combines origin-admissible cross-market attention with constrained spatial restrictions to significantly improve global volatility predictions across asynchronous markets, particularly at medium-to-long horizons, by effectively filtering spurious signals from closed exchanges.

Original authors: Xinlin Zhao, Haotian Qiao, Ziyao Lin

Published 2026-08-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinlin Zhao, Haotian Qiao, Ziyao Lin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Predicting how much stock prices will swing is one of the most critical tasks in finance. Investors, banks, and insurance companies rely on these forecasts to decide how much risk they can take, how to price complex contracts, and how to protect their savings. For decades, the standard tool for this job has been a method that looks at a single market's own history, assuming that today's volatility is largely a reflection of yesterday's, last week's, and last month's turbulence. This approach works well because stock markets tend to be sticky; once they start shaking, they often keep shaking for a while. However, the world is interconnected. When a major market in New York crashes, it often sends shockwaves to London, Tokyo, and Sydney. The challenge for forecasters has been figuring out how to use these global signals without getting confused by the fact that the world does not trade on a single, synchronized clock.

The problem is that international stock exchanges operate on different schedules. While the New York Stock Exchange is closed for a holiday, the Tokyo market might be open, or the London market might be closing early. If a computer model tries to look at all these markets at once, it faces a confusing gap: some data points are missing because the exchange was closed, not because the market was calm. If the model treats a closed market as having zero activity, it invents false information. If it simply ignores the closed days, it breaks the timeline and loses valuable context. This creates a fundamental puzzle: how can you build a model that learns from the rest of the world without accidentally peeking into the future or misinterpreting a holiday as a moment of peace?

A team of researchers set out to solve this puzzle by building a new forecasting system that respects these time differences while carefully deciding which global signals are actually useful. They did not try to replace the trusted, simple methods with a massive, complex artificial intelligence that guesses everything from scratch. Instead, they created a hybrid system that uses a proven, simple model as a solid foundation and then adds a very specific, disciplined layer of global information on top. Their goal was to see if looking at other countries helps, but only if that looking is done strictly within the rules of time.

The researchers tested their system using daily data from eight major stock indices around the world, including the United States, Germany, Japan, and China, covering a period from 2006 to 2022. This dataset included many different market conditions, from steady growth to severe financial crises. They compared their new method against the standard single-market approach and against several other complex models that try to learn from global data without strict rules. The results showed that their disciplined approach worked better than the standard method. By carefully filtering out the noise of closed markets and only using information that was actually available at the time of the prediction, their system reduced the error in its forecasts. This improvement was consistent across almost all the markets they studied, particularly when looking ahead one day, five days, or twenty-two days.

What made this system successful was not just its ability to look at other countries, but how it decided what to look at. The researchers built in a "gate" for each market that learned how much weight to give to global signals versus its own history. They found that for short-term predictions, like looking one day ahead, the system worked best when it relied heavily on the market's own recent history and only used a small amount of global information. However, as the prediction window stretched out to a week or a month, the system learned to rely more on the connections between different countries. This suggests that while a market's own momentum is the strongest driver for the next day, the influence of global trends becomes more important for understanding the longer-term picture.

The study also revealed that simply adding more complexity does not always lead to better results. When they tested a version of the system that tried to learn from global data without any strict rules or "gates," it performed worse than the simpler, disciplined version. The complex model tended to get confused by the noise in the data, mistaking random fluctuations for meaningful patterns. The researchers found that the best approach was to keep the core prediction anchored to a reliable, simple baseline and only allow the global information to make small, controlled adjustments. This approach prevented the model from overreacting to temporary noise while still capturing the genuine ways that markets influence one another.

One of the most significant findings was that the improvement from using global data was not uniform. The system did not become a perfect predictor for every single market on every single day. In some cases, the standard method remained just as good, and in a few specific instances, the new system was slightly less accurate. However, when looking at the average performance across all eight markets and all timeframes, the new system consistently produced more accurate forecasts than the standard method and outperformed other complex models. The statistical tests confirmed that these improvements were not just lucky guesses but were genuine gains in predictive power.

The researchers also discovered that the way they handled the missing data from closed markets was crucial. By treating a closed market as "missing" rather than "zero," and by ensuring the model never used information from the future, they avoided a common pitfall that plagues many forecasting attempts. This strict adherence to the rules of time meant that the model was learning from the past in a way that could actually be used to predict the future. The study concludes that in the noisy, chaotic world of finance, the most effective tools are often not the most complex ones, but rather those that are carefully constrained to respect the limits of available information. By combining a solid, simple foundation with a disciplined, rule-based look at the rest of the world, it is possible to see the future a little more clearly.

Technical Summary: Forecasting Global Volatility Across Asynchronous Markets

1. Problem Definition

The paper addresses the fundamental challenge of forecasting multivariate realized volatility across international equity markets where trading calendars are asynchronous. International exchanges operate on disparate holiday schedules, creating a "union calendar" where some markets are open while others are closed on any given date.

This asynchrony introduces two critical issues for forecasting:

Information-Set Violation: Traditional approaches often truncate non-common days (losing information) or impute missing data with zeros (injecting noise). Crucially, in multi-step forecasting, models risk inadvertently incorporating future information if the predictive set is not strictly bounded at the forecast origin ( $o = t - h$ ).
Overfitting in High-Dimensional Spaces: Recent literature employs complex machine learning architectures (e.g., STGNNs, Transformers) to capture non-linear spillovers. However, financial volatility has a low signal-to-noise ratio. Unconstrained neural networks in this environment are prone to overfitting transient transactional noise rather than learning robust economic linkages, often violating strict out-of-sample information boundaries through static, full-sample covariance matrices.

The central research question is not whether a complex neural network can outperform linear models unconditionally, but whether carefully regularized, origin-admissible cross-market information yields statistically significant incremental forecast accuracy beyond established volatility models.

2. Methodology: The PGA-Trans-HAR Framework

The authors propose PGA-Trans-HAR, a parsimonious hybrid forecasting system designed to test the value of constrained cross-market information. The architecture integrates a frozen, strong econometric baseline with a bounded neural correction, governed by four specific structural restrictions:

A. Forecast-Origin-Admissible Information Alignment

The system operates on a union calendar but strictly distinguishes between observed market-days and exchange closures.

Masking: An asymmetric mask ( $m_{t,n}$ ) ensures that closed exchanges do not transmit spurious zero-volatility signals as temporal or spatial sources, while still allowing them to act as destination queries.
Strict Origin Bounding: All inputs, including the neural window and the rolling econometric prior, are indexed by the forecast origin $o = t - h$ . No observation in the interval $(o, t]$ is included.

B. Constrained Cross-Market Information Channel

The model combines a data-driven spatial attention mechanism with a structural econometric prior:

Rolling Ridge-VAR/GFEVD Prior: A ridge-regularized Vector Autoregression (VAR) is estimated on a trailing historical window (up to 252 days) ending at the forecast origin. The Generalized Forecast-Error Variance Decomposition (GFEVD) converts this into a row-normalized matrix of directional predictive connectedness. This prior is refreshed every 20 forecast origins, ensuring no look-ahead bias.
Data-Driven Attention: A spatial self-attention mechanism learns dependencies directly from the masked input data.
Market-Specific Gating: A learned, time-invariant gate $g_n$ (specific to each destination market $n$ ) governs the convex allocation between the data-driven attention and the econometric prior. This gate is constant across dates within a fitted model, capturing persistent heterogeneity in how markets absorb foreign signals.

C. Benchmark-Centered Nonlinear Forecasting

Rather than replacing classical models, the system uses the Heterogeneous Autoregressive (HAR) model as a frozen anchor.

Bounded Residual Correction: The neural system predicts only a bounded residual correction in the inverse-softplus domain. The final forecast is:
$\hat{x}_{t,n} = \text{softplus}\left( \text{softplus}^{-1}(\hat{x}^{\text{HAR}}_{t,n}) + \alpha_n r_{b,n} \delta_{b,n} \right)$
where $\hat{x}^{\text{HAR}}$ is the frozen linear HAR forecast, and the neural term is strictly bounded by $\alpha_n$ and a sigmoid gate $r_{b,n}$ . This ensures positivity and guarantees that the neural component provides only incremental value.

D. Node-Balanced Optimization

The loss function includes a term penalizing severe relative deterioration in any single market, preventing the model from optimizing aggregate panel performance at the expense of specific markets.

3. Key Contributions

Methodological Rigor in Asynchronous Settings: The paper formalizes a strict protocol for handling asynchronous trading calendars, ensuring that all predictive inputs are strictly observable at the forecast origin, thereby eliminating look-ahead bias in both raw inputs and estimated relational graphs.
Parsimonious Hybrid Architecture: It demonstrates that a constrained hybrid system (Econometric Prior + Data-Driven Attention + HAR Anchor) can outperform both pure linear models and unconstrained deep learning architectures in low signal-to-noise environments.
Structural Ablation of Complexity: Through nested ablations, the paper isolates the specific value of architectural components, showing that flexibility is not monotonically beneficial and that structural regularization is essential for incremental gains.
Robust Statistical Evaluation: The study employs a comprehensive evaluation suite including 5-seed ensembles, HAC-adjusted Diebold–Mariano tests, and Model Confidence Sets (MCS) across daily, weekly, and monthly horizons.

4. Empirical Results

The study utilizes high-frequency data from eight major global indices (SPX, GDAXI, FCHI, FTSE, OMXSPI, N225, KS11, HSI) from 2006–2022.

Performance vs. HAR Baseline:
- Daily ( $h=1$ ) and Weekly ( $h=5$ ): PGA-Trans-HAR reduces both MSE and MAE across all eight markets relative to the univariate HAR benchmark.
- Monthly ( $h=22$ ): The constrained framework reduces errors for seven of the eight markets.
Performance vs. Broader Benchmarks:
- Daily: Achieves the lowest cross-market average MAE (though linear HAR-KS minimizes MSE).
- Weekly and Monthly: Achieves the lowest average MSE and MAE across the entire suite of linear and deep learning benchmarks (including VHAR, GNN-HAR, DCRNN-HAR, and Pure-ST-Transformer).
Statistical Significance:
- Diebold–Mariano tests confirm statistically significant improvements, particularly under absolute loss (MAE) at daily and weekly horizons.
- Model Confidence Sets (MCS) show the model is retained in the superior set for all markets at weekly and monthly horizons, though the daily horizon shows more competition with scalar-gated variants.
Ablation Insights:
- Horizon Dependence: The rolling econometric prior alone provides the most significant gains at the daily horizon.
- Gate Efficacy: The learned market-specific gate ( $g_n$ ) improves accuracy over a uniform scalar mixture ( $g=0.5$ ) at weekly and monthly horizons but degrades performance at the daily horizon. This suggests that short-term forecasting benefits from stronger, uniform shrinkage, while longer horizons benefit from differentiated cross-market allocations.
- Complexity: Unconstrained deep attention (Pure-ST-Transformer) fails to outperform HAR at short horizons, confirming that architectural flexibility alone is insufficient without economic regularization.

5. Significance and Claims

The paper claims a qualified affirmative conclusion: disciplined, origin-aligned cross-market information yields genuine incremental predictive value over the HAR anchor, with the most pronounced gains emerging at medium and long horizons where structural spillovers persist.

The authors emphasize that their contribution is not the superiority of deep learning over econometrics, but rather the value of constrained information. The results suggest that in financial forecasting environments with low signal-to-noise ratios:

Architectural flexibility is most effective when governed by economically interpretable restrictions (e.g., origin-admissible priors, bounded residuals).
The "black box" approach is less effective than a hybrid approach that anchors predictions in established volatility dynamics (HAR) and allows neural components to provide only bounded, incremental corrections.
The specific combination of a rolling GFEVD prior, asymmetric masking, and a market-specific gating mechanism successfully isolates the marginal predictive contribution of cross-market information without overfitting.

The paper explicitly avoids claims of universal model dominance, structural causal identification, or the ability to predict crisis regimes endogenously. Instead, it positions the framework as a robust method for extracting incremental accuracy from cross-market data while strictly respecting the boundaries of the available information set.