Hi folks,I am interested in measuring how well option-implied volatilities predict actual future realised volatilities in the forex markets.I use data from Bloomberg - my question: Is it correct to compare, say, the Bloomberg 3-month implied volatility data for Jan 1st with the Historic volatility on April 1st?Or in other words does the Bloomberg reported historic vola on April 1st concern the 3 months before that date (should be no, since there is already data available for today) and the implied vola for the 3 months after the reporting date?1) Here are my results for the EUR/USD exchange rate using the implied vola from Jan 1st with the historic Bloomberg vola from April 1st:2) And here are the results using both measures from the same reported day i.e. imp vola and historic vola from Jan 1st:Does Method 1) make sense and the lag between implied and historic vola is correct?Thanks very much indeed for your help!

[Removed double post]

Yes, but there is a potential source of confusion.Each day Bloomberg publishes an estimate of implied vol; that estimate is forward looking (i.e relates to the coming three months) and is statistically independent of prior estimates. Each day Bloomberg publishes an estimate of historical vol; obviously, since no one knows the future, they base that on the past three months (i.e. backward looking vol). What is important to understand is that these estimates made on successive days are not statistically independent of each other. In fact the estimates made on day n and day n+1 come from squaring and adding essentially the same set of 60 numbers (actually one number has been dropped and one number has been added at the other end, but the other 58 numbers in the middle are exactly the same). This accounts for the visual lag effect in your first chart. In fact, if you have a year of data, you can only extract 4 statistically independent estimates of 3-month vol, using 4 non-overlapping three month periods. These you can compare to 4 estimates of implied vol. To put it another way you cannot compare 250 daily estimates of RV with 250 (a year's) daily estimates of IV unless you take into account the statistical dependence of the RV's; even if you do take it into account, the fact remains that there are only 4 degrees of freedom per year of data that you use. Which will limit the statistical power of your study.So the simplest approach would be to compare 3-month implied volatility data for Jan 1st with the Historic volatility on April 1st, the IV for April 1st with the historical for July 1st, etc.HTH

Thanks for the input!When I look at the data then, the predictions of implied vola are slightly undershooting or am I seeing this wrongly? I.e. implied vola predicts say 20% for the next 3 months but this level is reached earlier by the historical vola.I thought I read that it should be the other way around or is the effect in my data more ambiguous?Thanks again!

Undershooting the changes -- that's what you see when you are predicting what amounts to the drift of a series dominated by noise -- in this case, the large vol. of vol. Volatility is modestly forecast by both the implied vol. and, alternatively, GARCH-type models, but most of the movement is due to unpredictable noise.

Construct a theoretical portfolio of long at-the-money options. Do they break even? Is their pattern of returns appropriately volatile?

Just to throw in my two cents...The market usually prices the implied volatility higher than the actual (historical) volatility (due to risk premium that the seller demands). In times of relatively calm markets the GARCH estimation and historical volatility should be pretty much inline with the implied vol, but in times of turmoil (like now, for example) the implied vols are priced much higher than the actual volatiltiy. This phenomenon is mostly noticable in the JPY crosses (USDJPY, AUDJPY, EURJPY...), as the volatility buyers are willing to pay a lot of money to protect themselves against chaotic movements.

If you are only constructing implied from ATM options, you have a skewed data set. You need to also factor in the information provided by OTM options (although you may assign it a lower weight). Many studies on implied vs historical miss this fact. See Canina and Figlewski and Taleb for references. There's also the issue of how much historical vol to use. Should you use the last three months to forecast the next three months or the last six months to forecast the next six months. Figlewski also covers this topic really well in his manuscript. pages.stern.nyu.edu/~sfiglews/Docs/Forecasting%20Volatility.pdf

That's great thanks!

QuoteOriginally posted by: frattyquantIf you are only constructing implied from ATM options, you have a skewed data set. You need to also factor in the information provided by OTM options (although you may assign it a lower weight). Many studies on implied vs historical miss this fact. See Canina and Figlewski and Taleb for references.If you buy ATM options and make money, nobody is going to take it away because you didn't own any OTM.