Deconstructing the VIX
Anytime we want to talk about “deconstructing” something, it helps to know what that something is. Here’s a good definition. Probably the key concept is that index options are sampled and calculations based on them become the VIX. This is also a good definition link because it includes the misnomer “investor fear gauge.”
I had been having a conversation with Bill Luby of VIX and More about this topic. My distaste for using the VIX can be summed up by its easy model-ability based on price action; if we already have price action in our model, VIX adds comparatively little useful information. VIX, regardless of actual construction methodology, can be described in terms of three contributors:
1) Actual, historical volatility
2) Some predictable measure of fear/greed resulting from recent price momentum
3) An error term.
NOW we might have something interesting. If we assume that a modeled VIX (based on volatility and momentum in the index) contains no useful information outside of that in the price data itself, then the deviation of actual from expected VIX could be said to contain “pure sentiment data” that corresponds to the pricing biases of market participants.
It’s a fairly simple process to work through, but time-consuming if one wants to optimize the formulae.
The first step is to date-match the data so that VIX and SPX data coincide; from 1/2/1990, I have three more days of SPX data than I do VIX data, probably because of different days the different exchanges were open – I discarded the three extra days in my data set, but you may choose to do something different.
The second step is to determine what “historical volatility” is. One could choose to use a standard deviation of the closing prices over some time period (what time period?), or use “typical prices” (average of Open/High/Low/Close? Ignore the Open?) over some time period, or use the Average True Range over some time period, etc. This choice could be done with single-variable analysis, or one could choose multiple definitions of “historical volatility” and run several multi-variate regressions in combination with the other variable …
Which is the third step, determine what measure of price momentum best predicts the level of fear/greed in the marketplace. This should probably be a timeframe that corresponds with a high level of mean reversion in the marketplace, so it could be a percentage MACD (PPO) of some length, a rate-of-change (ROC) calculation of the closes, or even! a 2-period RSI range.
Step four is to start running regressions. One could, of course, have run single-variable regressions to determine the highest correlation of VIX to various “historical volatility” measures and the optimum mean-reverting measure of price momentum, or one could merely run bunches of regressions, one against every possible pairing of measures, in a surfait of brute force calculations. Booyah!
Now that we have a good, bivariate regression formula that predicts VIX very well, using only historical volatility in price and recent momentum, we can compare this value to the actual VIX and see what the error term is.
Only when we have this “error term” do we have a true, unique measure of sentiment. VIX, in and of itself, is impure, i.e., it is mostly comprised of data that can be obtained from the movement of the SPX itself. If the VIX is high, does it really mean there’s inordinate fear in the market? I would suggest that the valid comparison is most likely “is the VIX high (low) compared to what I would expect given recent volatility and returns?” and not “is the VIX high (low) compared to historical or recent norms?” The “error term” between modeled and actual VIX is apt to give us better insight than just the VIX itself.
The penultimate step in the puzzle would be to compare this “error term” to various future returns on the SPX (2-, 5-, 10-day returns?) and find which ones were best predicted by the use of sentiment. The final step would be to compare the modeled result to other trading plans for the SPX to see if this work not only yielded something trade-able, but profitable enough to be worth the effort.


April 10th, 2007 at 9:11 pm
Great start, Bill. I’ll be waiting eagerly for the next chapter(s).
April 11th, 2007 at 2:32 am
Wow…yes, the error. I will be eagerly awaiting future installments on this subject.
April 11th, 2007 at 9:10 am
Nice work, Bill. I’m a big fan of Bill Luby’s work, and think the informal “Manhatten Project” you two have started on cracking the VIX is a great thing. Looking forward to the next installment.
best
David
April 11th, 2007 at 9:43 am
You can usually spot the error visually, at least on an intraday basis. My favorite setup is when SPX, DJI, or NDX are doing one thing mid-day and their corresponding volatility indexes are doing something “out-of-line”. This is a big hint at probable price action in the cash market later that afternoon.
What happens is that a big fund may know that it’s going to buy or sell a large market-moving chunk of securities on NYSE or NASDAQ later on that afternoon. So it starts accumulating leverage in the futures and options markets to take advantage of the big move to come. This action will affect options prices (hence options volatility apparent on the VIX). Then the time comes and the fund either buys or sells in the cash market causing volume to soar and a corresponding move in the futures/options markets. But the “smart money” has already been positioned. This is institutional manipulation at its finest and I have made a lot of money using this method.