Volatility is a feature that is present in any market, and so its analysis and forecast is something that can give you an advantage whatever market you are trading (and whether you build a strategy around it or you just use it as some sort of filter).

What is more, volatility as a process is (or *should* be) more easily forecastable than simple returns (although of course taking advantage of it is a little bit trickier).

The hot topic in the current volatility literature is the usage of High-Frequency data to get an estimate of a certain day volatility. This is the so-called Realized Volatility (RV).

Citing an article from Fulvio Corsi:

*“Importantly, Andersen, Bollerslev, Diebold, and Labys (2003) showed that direct **time series modeling of realized volatility strongly outperforms, in terms of outof-sample forecasting, the popular GARCH and stochastic volatility models.”*

RV is constructed from the sum of intraday returns, given a certain time sampling frequency. This differentiates this class of models from traditional GARCH-type of models, in that the latter you consider the volatility of a single day unobservable and instead calculate it as a function of the variance of daily returns over a certain periods.

For a certain day, we simply define the daily RV as the square root of the sum of squared logarithms price returns over the decided sampling number of minutes. It’s important to specify that in the literature, besides the use of squared returns, other regressors are considered too and according to the regressor used our RV time series will have different features. An example of these are absolute returns, called Realized Power Variation .

About the choice of the “optimal” sampling frequency: on one hand, the use of higher frequency is advisable because it allows to consider the sum of returns as an integral (for a number of time steps tending to infinity), and hence to have a “good” approximation of what’s called “integrated volatility”. On the other (more practical) hand, higher frequencies are affected by microstructure effects (as bid-ask bounce and order books dynamics in general). A way around this is to use models that account for this, or simply to use relatively lower frequency (20 or 30 minutes should do it).

This is the realized volatility of a synthetic Bunds Futures contract using absolute values and a sampling frequency of 20 minutes:

Once we have an estimate for the volatility on each day of some past history, we are ready to look for a model to forecast future realized volatility.

Andrea

Hi – tks for sharing – is it possible to get some code samples of this volatility article, to play with it?

Tks again

L.