Paper: arXiv 2308.01419
Authors: Chao Zhang, Xingyue Pu, Mihai Cucuringu, Xiaowen Dong
Abstract
We present a novel methodology for modeling and forecasting multivariate realized volatilities using customized graph neural networks to incorporate spillover effects across stocks. The proposed model offers the benefits of incorporating spillover effects from multi-hop neighbors, capturing nonlinear relationships, and flexible training with different loss functions. Our empirical findings provide compelling evidence that incorporating spillover effects from multi-hop neighbors alone does not yield a clear advantage in terms of predictive accuracy. However, modeling nonlinear spillover effects enhances the forecasting accuracy of realized volatilities, particularly for short-term horizons of up to one week. Moreover, our results consistently indicate that training with the Quasi-likelihood loss leads to substantial improvements in model performance compared to the commonly-used mean squared error. A comprehensive series of empirical evaluations in alternative settings confirm the robustness of our results.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper utilizes advanced graph neural network architectures with multi-layer setups and specialized quasi-likelihood loss functions, indicating significant mathematical complexity. It also demonstrates high empirical rigor through comprehensive out-of-sample backtesting across market conditions, alternative data splits, and robustness checks using realized volatility data.
Research Flowchart
flowchart TD
A["Research Goal: Forecast Multivariate Realized Volatility<br>with Spillover Effects"] --> B["Input: Multivariate Equity Volatility Data"]
B --> C{"Methodology: Customized GNN<br>with Multi-hop & Nonlinear Spillovers"}
C --> D["Training: Comparison of<br>Quasi-Likelihood vs MSE Loss"]
D --> E["Empirical Evaluation<br>in Alternative Settings"]
E --> F{"Key Findings/Outcomes"}
F --> F1["Nonlinear Spillovers<br>Improve Accuracy (1D-1W Horizon)"]
F --> F2["Quasi-Likelihood Loss<br>Significantly Outperforms MSE"]
F --> F3["Multi-hop Spillovers Alone<br>No Clear Advantage"]
F --> F4["Results Robust<br>Across Alternative Settings"]
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 GNN-based methodology (GNNHAR) for modeling and forecasting multivariate realized volatility with spillover effects.
Source sentence
In this article, we propose a novel methodology GNNHAR for modeling and forecasting RV, while taking into account volatility spillover effects in the U.S. equity market.
Assumes 6
- The input vector of N nodes is drawn from a multivariate Gaussian distribution. statistical
Source sentence
It assumes the input vector of N nodes is drawn from a multivariate Gaussian distribution N(0, Σ).
- Stocks among the DJIA components traded continuously throughout the entire sample period. data
Source sentence
we include only those stocks among the DJIA components that traded continuously throughout the entire period.
- Models are re‑calibrated monthly using a rolling window of the past 1000 days. model
Source sentence
All models are re-calibrated every month based on a rolling sample window of the past 1000 days
- Training uses 36‑month data and validation uses the most recent 12‑month data. model
Source sentence
Specifically, we use 36-month data for model training, and the recent 12-month data as the validation set to tune the hyperparameters
Show 2 more
- Adjacency matrix is built via Graphical LASSO to capture conditional independence of assets. model
Source sentence
Based on these compelling results, we adopt Graphical LASSO to establish the volatility graph in our study.
- The normalized adjacency matrix W = O^{-1/2} A O^{-1/2} is used to avoid numerical instabilities. model
Source sentence
where W = O^{-1/2} A O^{-1/2} is the normalized adjacency matrix, used to avoid numerical instabilities and exploding/vanishing gradients during the training phrase.
Data
- LOBSTER intraday data order-book commercial
Source sentence
The intraday data of Dow Jones Industrial Average (DJIA) components are obtained from the LOBSTER database.
- S&P100 components market-prices unknown
Source sentence
Additionally, for robustness checks, we consider a larger universe of S&P100 components.
Costs
Not applicable
Finds 5
- Incorporating spillover effects from multi‑hop neighbors alone does not provide a clear advantage in predictive accuracy. negative
Source sentence
Our empirical findings provide compelling evidence that incorporating spillover effects from multi‑hop neighbors alone does not yield a clear advantage in terms of predictive accuracy.
- Modeling nonlinear spillover effects improves forecasting accuracy of realized volatilities, especially for short‑term horizons up to one week. predictive-accuracy
Source sentence
However, modeling nonlinear spillover effects enhances the forecasting accuracy of realized volatilities, particularly for short‑term horizons of up to one week.
- Training with the QLIKE loss leads to substantial improvements in model performance compared to the commonly‑used MSE. performance
Source sentence
our results consistently indicate that training with the Quasi‑likelihood loss leads to substantial improvements in model performance compared to the commonly‑used mean squared error.
- The nonlinear GNNHAR1L model outperforms the linear baseline and avoids over‑fitting. performance
Source sentence
GNNHAR1L shows significant promise as a model for capturing nonlinearity, while avoiding the overfitting problem.
- Increasing the number of GNN layers generally degrades forecasting performance, though QLIKE‑trained models remain beneficial. performance
Source sentence
Generally, as we increase the number of layers in the GNNHAR models, their forecasting performance tends to decline. Nevertheless, we still observe the benefits of training models with the QLIKE loss function.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 44 related papers). Coverage is partial, so this is neither replication nor refutation.
2 papers 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.