Skip to content
All library documents

Modeling Event Effects and Outliers in Intraday Volatility

Article Quant Q&A · Author: Ana
Read the original

Summary

The question concerns estimating how a multi-day event affects volatility in one-minute equity returns, where extreme observations challenge normal-error ARCH and GARCH specifications. The response recommends matching the conditional error distribution to heavy tails or skewness, with Student-t, skewed-t, generalized error, and skewed generalized error distributions offered as examples. This lets a volatility model accommodate unusually large or asymmetric returns without assigning a separate dummy to each observation.

To estimate the event’s contribution, include a dummy for the event window in the volatility model. If alternative error distributions and the event indicator do not adequately capture the behavior, jump models are another possibility. The answer cautions that jump models can be difficult to calibrate and suggests treating them as a later option. It gives modeling guidance but no comparative results, diagnostics, or evidence that one specification will work across different assets and event types.

Key ideas

  • Heavy-tailed or skewed error distributions can better represent extreme intraday returns than a normal distribution.
  • An event-window indicator can be used to estimate the event’s effect on volatility.
  • A separate dummy for every outlier is not necessary in the suggested approach.
  • Jump models are an alternative when simpler distributional adjustments are inadequate, but calibration can be difficult.

Tags

EquitiesVolatilityEvent-drivenStatistics
Full text
# Events effect on intraday volatility and large outliers


# Events effect on intraday volatility and large outliers












I have an event that takes place over a period of a few days, and I want to estimate the effect it has on market volatility using intraday data with one minute frequency. The problem is, that e.g. GARCH, EGARCH and GJR are not able to account for the asymmetry in the distribution of the observations because there are several observations that lay well outside normal distribution (e.g. many of the stocks have 20+ 5 sigma returns during the sample period, majority of these during the event window).

Since I am interested in the effect that the aforementioned event has on volatility, I am trying to understand how to deal with these outliers. They obviously carry relevant information, and thus including dummy variables to account for every outlier would seem impractical and not very reasonable? My initial intuition for solving the problem was running whatever ARCH model gives the best fit and include dummy for the event period, but now I am starting to question whether this is the right method with the data I am working with after all.

All advice and references to articles I might've missed are greatly appreciated.

## Answer by Neeraj (score 2)

https://quant.stackexchange.com/a/24738

#### There are various alternatives to your problem:

- If you think that Normal distribution is not appropriate then you can use other distributions like t-distribution, skewed t-distribution, generalized error distribution, skewed GED etc. All these distributions are available in R (rugarch package) and Eviews too.

- As you are interesting in studying the impact of event on the volatility, the best approach is to use dummy variable for the event. You are not required to use dummy variable for the outliers. In this way, you can measure the impact of event on the volatility.

- If above two option does not work, You may also use jump models. There are vast literature on the jump models (just google it), but these models are difficult to calibrate. So, keep these models always your last options.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.