# What volatility estimator for continuous data and small time window?

I want to know which volatility estimator should I use for the following scenario:

I am implementing a market making bot and therefore I need to make estimations of the volatility of the price in the fashion of asking: What was the volatility of the price in the last couple of Minutes? (5-30 Minutes)

The data I've got available for the estimation is a set of all the prices in that time period at an interval of about 2-5 seconds, which are about 500 data point. And every time a new data point is added to the set, all data points that are older than the period of interest (5-30 Minutes) are removed from it.

Right now I use a basic estimator that calculates the variance of all the prices, however the problem with that is: the volatility oscillates way to much. I would expect the volatility to change slow and continuous over time.

First, you should use an exponential moving average, since the amount of state you need to keep is much smaller than for a simple moving average.

Second the well known estimator of volatility,

$$\hat{\sigma} = \sqrt{\frac{1}{n}\sum_{i=1}^n (x_i - \bar{x})^2}$$

is not very robust, since the squaring amplifies the contribution of outliers (which is why you are observing a very noisy volatility estimate - high frequency data has a lot of outliers).

Instead, consider using a mean absolute deviation estimate,

$$\hat{\sigma}_{MAD} = \frac{1}{n} \sum_{i=1}^n |x_i - \bar{x}|$$

which is more robust to outliers. You need to multiply this by a factor of $\sqrt{\pi/2}$ so that it matches the scale of the standard deviation estimator above.

The quantities $x_i$ should be the price differences from tick to tick, i.e.

$$x_i = p_i - p_{i-1}$$

or maybe the returns,

$$x_i = \frac{p_i}{p_{i-1}} - 1$$

You might want to consider thresholding the $x_i$ to some maximum value, say 5x or 10x the current volatility estimate, to reduce the impact of outliers even further.

• In the lecture I have about the market making algorithm I use, they propose the estimator by Garman and Klass. What do you think about that ? – flxh Sep 16 '16 at 10:17
• Garman Klass is a classical well known estimator, but perhaps not specifically developed for such short term data (was developed for daily OHLC, I believe). Bur probably OK to use. – noob2 Sep 16 '16 at 13:42
• From the lecture I talked about: "A second approach is to use the estimator proposed by Garman and Klass [60] (orig- inally devised for volatility across different days) adapted to the case of a time window of several minutes." Do you think Garman and Klass would fit even better here ? – flxh Sep 16 '16 at 15:24
• @Chris Taylor You say I need to calculate the moving average. Do you mean MA like a smoothed graph? Because in the formula you gave me the average has no index i so I assumed I need to calculate the average of all values over the whole time window?! – flxh Sep 16 '16 at 15:26
• No, you can compute a moving average of returns (though to be honest, this will be so close to zero most of the time that you might as well use $\bar{x} = 0$). – Chris Taylor Sep 16 '16 at 15:41