Higher Order Dynamic Network Linear Models for Covariance Forecasting
The paper proposes dynamic GNAR-HAR models that forecast realized covariances and volatilities, improving risk estimation and portfolio performance.
What it examines
This paper presents novel methods to forecast a portfolio’s covariance matrix by constructing dynamic correlation and volatility networks. Using network autoregressive (GNAR-HAR) models with higher-order neighbour interactions, the study aims to improve risk predictions and capture market behaviors, especially during volatile periods.
What it concludes
The results show that dynamic volatility networks with second-order correlation interactions yield better forecasting accuracy and higher portfolio performance, particularly in turbulent markets. This research can be applied to portfolio management, risk control, and asset allocation. Future work will refine graph construction and parameter estimation to further improve financial forecasts.
Evidence objects
Researchers present a forecasting approach merging network-based models with the HAR method, using dynamic volatility networks and higher-order interactions via data-driven CoC graphs to capture asset interdependencies and market shifts.
key_findings bullet 1 · key_findings · validation V0
Integrating dynamic volatility networks, the model reduces forecasting errors and improves portfolio performance amid market turmoil, yielding smoother parameter estimates, higher Sharpe ratios, and lower turnover via risk management techniques.
key_findings bullet 2 · key_findings · validation V0
Using rolling window estimation, winsorisation, and Graphical Lasso, the study builds sparse volatility networks and tests graph structures. It reveals sensitivity to network construction and update frequency, prompting further research.
key_findings bullet 3 · key_findings · validation V0
This paper introduces an innovative framework integrating dynamic network models and time-varying graphs to forecast covariance matrices, addressing portfolio risk management and prediction challenges. Its novel application of higher-order ($GNAR$) interactions and Graphical Lasso enhances volatility and correlation dynamics, offering compelling, original, and impactful advancements over traditional methods in modeling.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … Given a universe of N assets and access to N high-frequency time series price returns, we are interested in forecasting the next day return covariance matrix. …
Source row: 1036 · abstract type: snippet