@Smokeshow

Here is the big question, in my view: if asset volatility risk ranges come from volatility of volatility, how would one apply this to an individual equity that does not have a vol of vol product such as VVIX?

IMHO there are two ways to do this:

1. Instead use realized volatility of recent implied volatility readings
2. Use ratio back spreads on an individual stock as a proxy for Ivol of Ivol.

Regarding (2), we need to think a little bit about what vol of vol really is. Vol of vol is the sensitivity of out of the money options in relation to those at the money. It can be called gamma of gamma or the 4th moment.

Ratio back spreads were the traditional way (before vol of vol products on indices) to capture (be long) vol of vol. This implied selling x ATM call and buying x OTM calls (say sell one call, buy 2), in a manner that hedges delta. This isn’t as clean as buying a vol of vol product, as lower moments (ie gamma) can take over.

The problem with using (2) for a vol of vol proxy on a stock to use in the RR’s is this: one would have to maintain a delta-neutral positioning to actually approximate vol of vol. This means that each day, one would have to calculate implied vol of implied vol on a stock by measuring the delta-hedged vega of the net position of a short ATM call with long OOTM calls, say 30 days out (or something close). It’s pretty messy, but I think doable with the right data streams.

I have a feeling that on gold, SPY etc, Keith uses either the Vol of Vol index (VVIX) on the S&P, or simply looks at the options pricing (implied vols) on something like GVZ.

On individual stocks, I think he is reverting to either #1 or #2. I am guessing it is #2, as its a better approximation of the vol of vol compared to Rvol of Ivol.

If I had to guess…

He’s not doing (2). It’s a little complex and sort of outdated.

This thread has done a great job talking about the concepts that HE refers to consistently. There’s one that he always talks about that is worth talking about: Bayesian inference.

I have a strong feeling this is being applied somehow into the RRs. I’ve tried a number of deconstruction methods and can get close, but there are changes that are confusing. The Hurst stuff – I think it may play into it, but on a one day basis (ie RR), Hurst doesn’t do much for you.

I’ve tried adjusting the IVOL profile based on vol of vol, and it doesn’t really move the needle enough compared to what he has put out there. IE he will have 18 point increases in risk ranges on SPX when IVOL/VIX actually decreases, and Vol of Vol is flat. There is something else going on here.

If I had to guess, he’s using Bayesian priors based on some vol calculation, and the formula is either dropping meaningful priors (periods), or changing the Bayesian prior vs posterior mix weighting (basically a base case distribution assumption combined with a more dynamic assumption. This goes hand in hand with ‘vol clustering’. As well as ‘regimes of vol’.

I am guessing when Ivol or vol of vol breaches some level (regime), either the posterior distribution changes dramatically (power law scaling), or a significant rewighting of prior/posterior occurs.

KM always talks about Bayesian processes. This would be the thinking to explain how some others have mentioned that the two ends of the risk range can change non-uniformly.

If I had to guess, his Excel based model that he made at Magnetar (or even 3-7 years ago) was very, very different than what he uses now, with teams of quants and coders. Similar basic principle (probabilistic distributions), but lots of bells and whistles that didn’t exist before.

I think we have enough clues to suggest that when the IV data is available, he uses it. For instance, he talks about RRs moving when GVZ or VIX changes significantly.

In the context of the recent and current market, this makes a lot of sense, as derivatives and market structure (option gamma in particular) have been major drivers of short term returns. Some of this would be “somewhat” accounted for in option pricing skew. But given what has happened the last year (dealer short gamma being so influential), I wouldn’t be surprised if he has incorporated some sort of gamma calculation. Take a look at what SpotGamma.com is doing to flesh out important levels by scraping options data.

In markets with very low option OI (ie BTC), I think he uses some various bootstrap techniques to approximate IV.

In short, I think he has supplemented/changed his RR inputs over time. After all, he has teams of quants and programmers on payroll. It’s hard to imagine a situation where he isn’t pulling this horsepower into his main short-term risk product. I am guessing that 10-15 years ago, it was not much different than many of the stochastic oscillators that one can find on Tradingview. Given how much he talks about Hurst, Bayes (which is definitely used in the GIP model), and implied vols, I really can’t see RRs being solely based on RVs.

my 2c…

I saw that, and I completely agree with you.

On Monday’s “The Call” Keith was discussing AMWL.  He mentioned that he did not have enough data to run a trade range on Monday but he would have enough data after Tuesday’s close to run the trade range on Wednesday, AMWL was listed on 9/17/2020.  That tells me that he is using either 18 or 19 days (depending on whether it is close to close rate of change or if it is the data at the close).

In that last couple of weeks we have, in my opinion, confirmed that Keith is using option data, probably vol or vol of vol, and he is using 18 or 19 days worth of data.

Here’s a thought that may address the risk ranges “shape” changing when other obvious things don’t (like the VIX vs SPX):

It’s possible he is using something such as implied call vol and implied put vol for the ranges.

Say what?

Calls and puts can have different implied vols (prices). This is a function of the market wanting to own more of one or the either, which changes price. Now why does this matter? Numerous studies have indicated that implied vol options prices have information not found in share prices.  (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2695145).

Bloomberg has a field in their API for call and put ivol. They are presumably taking a series of options (both puts and calls), and are calculating the Ivol in a similar way as in either the VIX (here: https://www.macroption.com/vix-calculation/) or weighting by OI maybe.

Why would this be useful? Let’s say you calculated call IVOL at 25 and put IVOL at 35. That is an annualized 1SD expected move. If we agree that there is data embedded in options (I do), and that’s what we’re trying to capture, we can make that into a risk range. But we need to pick the number of days. If the average days to expiration of the options that are being used in the IVOL calcs is 30 days. Remember that annualized IVOL #s use 365 day calendars (not 252).

call IVOL = 25% * √(1/365) x √(2/π) = 1.04% ‘implied average daily move’.

put IVOL = 40% * √(1/365) x √(2/π) = 1.67% ‘implied average daily move’.

Important: (a)1 SD is not the same as average daily move. I use ADM instead of SD as options are priced using mean outcomes, and people replicate hedges daily. (b) I am using 365 days as that is the standard as how vanilla options are priced. RV uses 252, and some some OTC vol swaps do use 252.

So, you want to think of a risk range in the context of some # of days? Let’s pick 18 days for fun.

call IVOL = 25% * √(18/365) x √(2/π) = 4.43% ‘implied average 18 day move’.

put IVOL = 40% * √(18/365) x √(2/π) = 7.09% ‘implied average 18 day move’.

Now, we need not worry about a MA for this calc, as we are using implieds. On realized, we need to think about the mean that is going into the calc of the SD/ADM.

How do I test this? Well I don’t know the series of options IVOL assumptions to be used. Maybe it’s the way Bloomberg does it, maybe not, no idea. So, let’s say I were to use AAPL Ivols over the next 30 days. (there isn’t much difference in AAPL between call and put IVOL).

(this method above is definitely wrong, I should be using more strikes and adjusting to variance contribution, but here ya go)

call IVOL = 44.8% * √(18/365) x √(2/π) = 7.94% ‘implied average 18 day move’.

put IVOL = 45.3% * √(18/365) x √(2/π) = 8.03% ‘implied average 18 day move’.

AAPL closed Oct. 15, 2020 at $119.02. That would give it a range based on my method of $109.90 – $128.47. Not bad eh? I’ll note that there was very little difference between put and call IVOL, so it wasn’t a great candidate.

One point regarding vol of vol. If you use a wider array of option strikes (mainly if you start incorporating OOTM strikes in the call IVOL and put IVOL strikes, unlike what I did), you can start incorporating some other embedded information. Skew is valuable data that may change the risk range (particularly to the downside). This would be weighted to the contribution to that expiration’s overall variance. This would also pick up some gamma pricing and vol of vol data.  In fact, certain option strategies like this one (https://www.fidelity.com/learning-center/investment-products/options/options-strategy-guide/1×2-ratio-volatility-spread-calls) expose you to vol of vol because of the structure and the OOTM calls that you are long.