Dynamic partial (co)variance forecasting model

Abstract

In this study, we propose a dynamic partial (co)variance forecasting model (DPCFM) by introducing a dynamic model averaging (DMA) approach into a partial (co)variance forecasting model. The dynamic partial (co)variance forecasting model considers the time-varying property of the model's parameters and optimal threshold combinations used to construct partial (co)variance. Our empirical results suggest that in both variance and covariance cases, the dynamic partial variance forecasting model can generate more accurate forecasts than an individual partial (co)variance forecasting model in both the statistical and economic sense. The superiority of the dynamic partial (co)variance forecasting model is robust to various forecast horizons.

Keywords:

• Partial (co)variance forecasting
• Realized variance
• Dynamic model averaging

JEL Classifications:

• C58
• C22
• G14
• G13

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplemental data

Supplemental data for this article can be accessed online at http://dx.doi.org/10.1080/14697688.2024.2342896.

Data availability statement

The data that support the findings of this study are available from Weisheng Database. Restrictions apply to the availability of these data, which were used under license for this study.

Notes

1 The sample data, comprising three random subsamples (January 1, 2013–June 30, 2016; July 1, 2016–June 30, 2019; and July 1, 2019–December 31, 2021), demonstrate the inconsistency in the predictive efficacy of realized partial variance forecasting. This inconsistency supports the research presented in this study.

2 In addition to the MSE loss function, We also present the results based on MSD and QLIKE loss function in supplementary material.

3 We use the confidence level of 90%, which allows us to exclude a model with a p-value smaller than 0.1. In other words, the forecasts of this model are significantly less accurate than other models in MCS. In this study, we use the SQ statistic to calculate the MCS p-value through 100 000 stationary bootstraps with a block length of 25.

Additional information

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.