Conditional Volatility Persistence and Realized Volatility Asymmetry: Evidence from the Chinese Stock Markets

ABSTRACT

This study proposes that the overall state of the market, as captured by daily return and volatility, is an important determinant of volatility persistence. By utilizing the realized variance (RV) measure, this paper shows that daily time-varying volatility persistence increases with return but decreases with volatility. Negative returns increase volatility persistence more than positive returns. The dependence of volatility persistence on state variables is termed “conditional volatility persistence”. This study finds that conditional volatility persistence is the dominant channel linking changing market states to future volatility and the model which calibrates future-RV conditionally on market states performs better statistically and economically.

KEY WORDS:

• conditional volatility persistence
• economic value
• realized variance
• volatility asymmetry
• volatility forecasting

JEL Classification:

• G12
• G14
• C13
• C22

Supplementary material

Supplemental data for this article can be accessed on the here.

Notes

1. For example, Wang and Yang (2018) propose three potential mechanisms that would probably lead to volatility persistence, including the endogenous/exogenous information flow, the persistence characteristics of information process, and uninformed trading, such as portfolio adjustments to lagged information shocks.

2. The annualized volatility is calculated as the squared root of daily RV multiplied by 250, then divided by 100.

3. The Chinese stock market turbulence began with the popping of the stock market bubble on 12 June 2015 and ended in early February 2016. A third of the value of A-shares on the SSE was lost within one month of the event.

4. Corsi and Reno (2012) demonstrate that the return impact on future volatility is also highly persistent and propose a HAR structure for returns to capture their heterogeneous effects. Accordingly, the weekly (r_i,t,W) and monthly returns (r_i,t,M) are added to the benchmark HAR model, which are similarly defined as (RV_i,t,W) and (RV_i,t,M).

5. For a detailed demonstration of the Shapley-Owen R² method, please refer to Appendix of Lahaye and Neely (2018).

6. To adjust for seasonality, we regress the market state variables on their own lagged values, time dummies (i.e., weekdays and months), linear and quadratic time trends, and then use the residuals as the seasonally-adjusted time series in the following regressions.

7. However, as explained in Bollerslev et al. (2018), the same setup readily extends to allow for time-varying Sharpe ratios as long as the temporal variation of Sharpe ratio is independent of the conditional risk.

8. The number of 10% volatility level is in the right ballpark for many target volatility funds, according to Bollerslev et al. (2018).

9. For a more detailed introduction to the derivation of the expression of average realized utility per unit of wealth (UoW), please refer to Bollerslev et al. (2018).

10. Following Wang and Yang (2018), an “insanity filter” is used to replace a negative RV forecast with the lowest RV in the rolling window.

11. For example, Patton and Sheppard (2015) show that the semi-variance HAR model has better RV forecasts than the standard HAR and a RV-based GJR-GARCH model.

12. In fact, shareholder transaction costs are usually zero for index mutual funds. With ETFs, shareholder transaction costs can be broken down into commissions and bid-ask spreads, which are generally very low as well.

13. In fact, the utility benefits of the CVP model are substantially larger if the SR and/or risk target are larger . For instance, if the SR is 0.4 and the annualized risk target is 15% (corresponding to a risk aversion of 2.7), then the relative utility benefits increase to 0.92% per year.