Thread by @HauVolatility

& Let the radio silence end. As you may have seen, we’ve spent the past couple of months researching, modeling, and building a new ‘ecosystem’ to track market gamma for individual stocks and SPX.

A quick explanation (and history):

Early models solved this problem by making assumptions about open interest.

The first / most popular OI assumption is just assuming all calls are sold and all puts are bought. Surprisingly, some people still use this assumption today to measure gamma.

This is frightening.

This assumption works *sometimes* on $SPX. Here is a plot of SPX '15-'16 (x axis = gamma exposure, y = 1-day return) & another with daily data from 2017-March 2020.

Nice.

A fan like pattern is what we’d expect a good model to produce (lower vol in high gamma and vice verse)

Sadly, open interest assumptions almost *never* work on individual stocks. Just imagine making an assumption about all calls in a single name like $TSLA indefinitely...

Not good. Check out $JPM and $GOOGL below.

Again, a good model would produce a fan shape.

& Yes, we tried tweaking all sorts of models to fit open interest for stocks.

I.e. X% of ATM calls are sold, X% OTM are bought, puts are never bought, etc. None of them worked, and the ones that did only looked nice due to extreme overfitting bias.

Checkpoint: Gamma models that rely on assumptions about OI are borderline useless in practice- ESPECIALLY on individual stocks. Even with $SPX assumptions, how do you know when you're in a 2011-2012 and when you're in a 2017-2020 before it's too late?

This is what compelled us to create a gamma model without assuming ANYTHING about open interest. For that, you’d have to *probabilistically* assign every options trade as a buy or a sell (look through EVERY transaction and get INSIDE open interest).

This takes us to Spring 2020.

Our first model established a gamma metric based on "flagging" trades probabilistically as a buy or a sell or a crossed/paired order based on the live order book.

It was a HUGE improvement to open interest models, especially on single names, but had minor kinks of its own...

Such as:

Trade reporting on the tape later than actual fills, pesky orders designed to throw off other market makers by faking liquidity, large spreads on names etc.

The new additions to the system, the model that kept us off Twitter for awhile, ironed out those wrinkles.

New system: a hybrid model that uses the live order book AND the market maker’s implied volatility surface to flag trades as ‘buys’ or ‘sells’ based on where they fill. You may be able to hide behind tricky quoting on the order book, but you can’t hide your vol surface from us!

A couple of custom Python packages and the torture of many academic papers later, you can track the quoting regardless of what's sitting on the top of the order book & have a robust model to trades that are filling "late" or "deceivingly".

Of course with this, you can produce output like what’s shown below: Flag millions of trades everyday as buys (green bars) or sells (red), calculate and track their IV and gamma, and maintain a very, *very* large database that tracks dealer’s inventory day in and day out.

Cheers