Story
Multivariate GARCH and portfolio variance prediction: A forecast reconciliation perspective
Key takeaway
Researchers found combining different forecasting models can improve predictions of portfolio risk, which helps investors better manage financial risk.
Quick Explainer
The work presents a novel approach to improve portfolio variance forecasting by combining the strengths of univariate and multivariate GARCH models. The key idea is to reconcile the forecasts from these two modeling approaches, leveraging their complementary perspectives on asset interdependence. The method first fits a univariate GARCH model on the portfolio returns to obtain a baseline forecast, then combines this with a bottom-up forecast derived from a multivariate GARCH model on the individual asset returns. This reconciliation process enhances the accuracy of the final portfolio variance prediction, even when the multivariate model is misspecified or the covariance proxy is noisy. The authors demonstrate the effectiveness of this technique through simulation studies and an empirical application.
Deep Dive
Technical Deep Dive: Multivariate GARCH and Portfolio Variance Prediction
Overview
This work presents a novel approach for improving the prediction of portfolio variance by combining univariate and multivariate GARCH models using forecast reconciliation techniques. The key contributions are:
- Demonstrating that forecast reconciliation can enhance the accuracy of portfolio variance prediction compared to using univariate or multivariate GARCH models alone.
- Showing that multivariate GARCH models can be susceptible to performance deterioration under model misspecification, and that forecast reconciliation can help mitigate this issue.
- Highlighting the crucial role of using less noisy proxies for the covariance matrix when contrasting forecasting models and approaches.
Methodology
The authors propose a forecast reconciliation approach to combine the portfolio variance forecast from a univariate GARCH model on the portfolio returns with the bottom-up forecast obtained from a multivariate GARCH (MGARCH) model on the individual asset returns.
The key steps are:
- Fit a univariate GARCH(1,1) model to the portfolio returns to obtain a "base" portfolio variance forecast.
- Fit a MGARCH model (either DCC-GARCH or Scalar BEKK) to the individual asset returns to obtain forecasts of the asset variances and covariances.
- Combine the univariate and multivariate forecasts using optimal linear reconciliation to obtain the "reconciled" portfolio variance forecast.
The authors also introduce two novel reconciliation strategies to ensure the reconciled correlation matrix satisfies the mathematical properties of a valid correlation matrix.
Data & Experimental Setup
The authors conduct an extensive simulation study to evaluate the performance of their proposed approach. They consider:
- Two data generating processes (DGPs): DCC-GARCH and Scalar BEKK, with and without interdependence between asset variances.
- Three sample sizes (T=500, 1000, 2000) and two portfolio weight schemes (equally-weighted, randomly-weighted).
- Three MGARCH models for estimation: DCC-GARCH, Scalar BEKK, and EDCC-GARCH (an extension of DCC-GARCH with interdependence).
They also examine the impact of using a noisy proxy for the true covariance matrix, as is common in empirical applications.
Results
The key findings from the simulation study are:
- When the MGARCH model is correctly specified, forecast reconciliation consistently improves upon the univariate and bottom-up approaches.
- Under model misspecification, forecast reconciliation still outperforms the individual models, even when the bottom-up approach performs poorly.
- The presence of a noisy covariance proxy can mask the benefits of accounting for interdependence, making misspecified models perform similarly to correctly specified ones.
- Increasing the number of assets from 9 to 24 amplifies estimation risk and misspecification costs, keeping the univariate baseline competitive.
The authors also provide an empirical application using daily returns on 28 Dow Jones Industrial Average constituents, which corroborates the simulation results.
Interpretation
The authors demonstrate that forecast reconciliation is an effective tool for improving portfolio variance prediction, as it can leverage the complementary strengths of univariate and multivariate modeling approaches. The results highlight the importance of properly accounting for asset interdependence, as well as the sensitivity of MGARCH models to model misspecification and noisy covariance proxies.
Limitations & Uncertainties
- The simulation study is limited to specific GARCH model specifications and does not explore the full range of MGARCH model types.
- The empirical application only considers a single dataset, and the findings may not generalize to other market conditions or asset universes.
- The authors do not explore the implications of their approach for risk management applications beyond variance forecasting, such as value-at-risk or expected shortfall.
What Comes Next
Future research could extend this work in several directions:
- Investigate the performance of forecast reconciliation with a broader set of MGARCH models, including more flexible specifications.
- Examine the application of reconciliation techniques to other risk measures beyond portfolio variance, such as conditional value-at-risk.
- Explore the integration of forecast reconciliation into dynamic portfolio optimization and risk management frameworks.
