Hi there,I have been calculating volatilities in a fairly simplistic manner?I take the return of the underlying:I then calculate the standard deviations of these returns using a non-overlapping 2 day window.Non when I calculate the acf for these volatilties I see quite strange patterns in the acf with the first three all generally being significant:The first lag the largest and always negativeThe other two much smaller but still significant sometimes positive and sometimes negative depending on the length of window used.Now when I test the underlying process for a unit root I find that it has a unit root.Therefore when I calculate the returns I remove the unit root.However when I further calculate the standard deviations I am effectively differencing again thus over differencing and so introducing an MA error into the process which is what I think is causing the ACF patterns.Assuming that is the case how can I control for/eliminate these ACFs which are being generated by my data processing techniques?I just want to produce a time series for volatility that reflects the true volatility dynamics.

QuoteI just want to produce a time series for volatility that reflects the true volatility dynamics.don't you want to build a Garch-family volatility model?anyway, you might be interested in E.Sinclair's book Volatility Trading (i'm not affiliated) . The author gives formulas, explains and compares several volatility definitions.

Thanks for this

If you have a negative autocorrelation at lag one with financial prices, it's usually due to mismeasured prices. If you have a positive error in the price at today's close, you will have a positive error in today's return and a negative error in tomorrow's (you use two day returns, but the principle is the same). Depending on what kind of prices you have, the error could be a stale price (maybe the last trade was a 1 PM for a market than closes at 4 PM), or some days the last trade is instigated by the buyer and other days by the seller, or just an error. Any autocorrelation at lag one induces a positive autocorrelation at lag 2. If you are adjusting for this, it is not surprising to see positive and negative significant autocorrelations at lags two and three.There are statistical techniques for cleaning the data, but in my opinion they are usually harmful. The best solution is to get additional information that allows you to remove the autocorrelation without reference to the data series itself. It can be as simple as cleaning the data.