# Volatility Modelling negative GJR-GARCH-X coefficient

I have estimated GARCH and GJR-GARCH with several exogenous variables. Some of the exogenous variables have negative coefficients that are statistically significant. For instance, I can write my GJR-GARCH estimate as:

$$h_t = 0.213 + 0.011 u_{t-1}^2 + 0.847 h_{t-1} + 0.196 u^2_{t-1}I_{t-1<0} - 0.026 X_{t-1}$$

Where $$u_{t-1}^2$$ is the lagged ARCH term, $$I_t$$ is the dummy variable that models the asymmetric leverage effect and $$X_{t-1}$$ is the exogenous variable.

My main concern is that the coefficient of $$X_{t-1}$$ is negative. How can I verify that my volatility is not negative? Is just a plot of the conditional variance over the in-sample period is enough?

I have tried to estimate the GJR-GARCH-X model, however, I obtained some negative coefficients. I am not sure how to verify that these negative coefficients will cause the volatility to be negative.

### Positivity of GARCH-X models:

When it comes to GARCH models with exogenous regressors, it is more "tricky" to ensure positivity of the model, due to the unspecified model-dynamics of your exogenous regressor (moreover exogenous regressors can have drastically different characteristics). Without any model-specification on the exogenous regressor $$X_{t-1}$$, it is common to restrict the parameter-space of the model in order to ensure positivity.

For simplicity, let us vaguely define the GJR-GARCH(1,1)-X model with demeaned returns, $$r_t$$: \begin{align*} r_t \vert \mathcal{F}_{t-1} &= \varepsilon_t\\ \varepsilon_t &= \sigma_t \cdot z_t\\ \sigma^2_t &= \omega + \alpha \varepsilon_{t-1}^2 + \beta \sigma_{t-1}^2 + \gamma I_{t-1} \varepsilon_{t-1}^2 + \nu X_{t-1}, \end{align*} where $$z_t \overset{iid}{\sim} D(0,1)$$ is a standardized distribution and

$$I_{t-1} =\begin{cases} 1 & \text{if } \varepsilon_{t-1} < 0 \\ 0 & \text{if } \varepsilon_{t-1} \geq 0 \end{cases}.$$

I have detailed some of my observations:

• When working under the GJR-GARCH(1,1) positivity is satisfied when we impose $$\omega, \beta,\alpha > 0$$ and $$\alpha + \gamma > 0$$. The latter condition is a broader statement than imposing $$\alpha, \gamma >0$$, since we can allow one of the parameters to become negative (in this case, $$\gamma$$). I have made an in-depth answer detailing the GJR-GARCH(1,1) model, positivity, covariance stationarity and economical interpretations of the parameter estimates.

• When working under the GJR-GARCH(1,1)-X we can further ensure positivity by additionally restricting the exogenous regressor such that $$\omega + \nu X_{t-1} \geq 0$$. Here, we allow $$\nu$$ to vary freely (since $$\omega > 0$$) as long as the above condition is satisfied. The primary motivation for the extra restriction, comes from observing the unconditional variance (calculated under assumed covariance stationarity) of the GJR-GARCH-X model:

$$\begin{equation} \mathbb{V}ar(r_t) = \frac{\omega + \nu \mathbb{E}\left[X_{t-1}\right]}{1 - \alpha - \beta - \kappa \gamma}. \end{equation}$$ In order to ensure non-negative unconditional variance of the return process, we specifically need $$\omega + \nu \mathbb{E}\left[X_{t-1}\right] \geq 0$$, which is satisfied when imposing $$\omega + \nu X_{t-1} \geq 0$$ for all $$t$$.

In conclusion, imposing $$\omega, \beta, \alpha > 0$$, $$\alpha + \gamma > 0$$ and $$\omega + \nu X_{t-1}>0$$, ensures that you obtain non-negative volatility estimates.

When $$X_{t-1}$$ is strictly positive, it is common to let $$\nu \geq 0$$. This is also emphasized in the article of Han, H. (2015) that investigates asymptotic results of the GARCH-X model when $$X_{t-1}$$ follows a fractionally integrated process. In general, it is common in academia to assume a functional form on the exogenous regressor, (see for instance (C1 - C4) in this paper p.699). This is also done in the Realized GARCH model that incorporates intraday data to procure better forecasts (I have detailed this model here, if you're interested). I hope this helps.

• This is very helpful response. However, I have couple of question to ask and will be grateful if you can answer for me: 1- My exogenous variable $X_{t}$ is a news sentiment variable that takes negative values with negative news and positive with positive news i.e. it is not strictly positive. My second question comes in a different comment due to the number of characters restrictions. Sep 9, 2022 at 18:40
• 2- I am using rugarch package on R for my model, this is my code: ug_spec_x <- ugarchspec(mean.model = list(armaOrder=c(0,0)), variance.model = list(model="sGARCH", garchOrder = c(1,1), external.regressors = matrix(C\$all.news)) ,distribution.model = "std" ) setbounds(ug_spec_x) <- list(vxreg1=c(-1,1)) ugfit_x <- ugarchfit(spec = ug_spec_x, data = C\$GSPC.Adjusted, solver= "solnp") How can I impose the restriction you advised me in my code ? i.e. $\omega + \nu X_{t-1} \geq 0$ Sep 9, 2022 at 18:41
• Thank you Pleb! I will wait to hear back from you Sep 10, 2022 at 0:08
• @Moataz Most of the papers I have been skim-reading works with a positive exogenous covariate. In that regard, I would transform my exogenous regressor into a GJR-type equation, $\nu I_{\{{X_{t-1}<0}\}} \varepsilon_{t-1}^2$. This will intuitively, give you an understanding on how negative shocks in your sentimental news covariate, affects future volatility of returns. This is very similar to the original GJR-term. Therefore, be aware that bad news are already "incorporated" into past squared returns and in such case, the original GJR-term will likely catch most of the asymmetrical response.
– Pleb
Sep 10, 2022 at 10:11
• Hi Pleb, thank you for your comemnt and your suggestion. If I followed your suggestion, then my unconditional variance will be $\frac{\omega}{1- \alpha - \beta - \frac{\gamma}{2} - \nu }$ ? Sep 10, 2022 at 12:30