Sunday, May 14, 2006

Spike in the VIX: Is There an Edge?

The VIX is a measure of option volatility, so when it spikes higher--as has happened in the past two trading sessions--it's a sign that market makers are pricing options for further movement. This typically occurs following market declines, which leads to the notion that an elevated VIX reveals trader fear, while a low VIX suggests complacency.

On Friday, the VIX closed 18% above its 20 day average--a substantial spike. Since March, 2003 (N = 804 trading days), we've had 32 occasions in which the VIX has closed 15% or more above its 20 day average. Two days later, SPY has averaged a gain of .58% (21 up, 11 down). That is much stronger than the average two-day gain of .12% (437 up, 367 down) for the sample overall.

This is consistent with Larry Connors' research, which suggests that VIX elevations lead to superior returns in the near term.


Barry Ritholtz said...

Ahhh, but Doc -- the period you referenced runs from a market bottom to present. Any thesis needs to be tested under various market conditions, and Connors VIX idea only looks at one cyclical Bull market -- and not a full cycle.

How did this VIX indicator do in 2000?

That doesnt mean this won't work here -- but I wonder how it will do when this cycle ends . . .

Brett Steenbarger, Ph.D. said...

Hi Barry,

I greatly enjoy your writing on the Real Money/ sites and in your Big Picture blog. Your point--and your caution--are well taken and touch upon an important different between quants who do non-linear historical modeling and quants who develop mechanical trading system ideas.

The developer of a mechanical system is, for the most part, looking for relationships that will hold up across a variety of market conditions. That is why the systems (ideally) will be tested over a sufficiently long period (and validated over a separate time period) to incorporate bull, bear, and flat market conditions.

The non-linear modeler is looking for relationships that are specific to market conditions similar to those at the present time. The nearest neighbor approach, for instance, only pulls out historical market periods similar to recent history to develop forecasts.

The great caveat of the non-linear approach is that it rests on the assumption that the next time period will be similar to the recent past being analyzed. In statistical terms, you're assuming that the next market days are drawn from the same sample (generated by the same underlying process) as the previous ones. At times when market cycles change, that assumption becomes tenuous.

Of course, the other way you could look at it is that when the performance of non-linear models degrade, this itself becomes an empirical criterion for identifying changing market cycles.

Readers interested in non-linear historical modeling might check out the Salford Systems site ( Meanwhile, I think a look to see how indicators such as the Relative VIX performed in 2000-2002 would be instructive.

Thanks for the prod!


Gammaman2k said...

I think you are being very selective in your data. Firstly the SPX made a low in March 2003 around 790 if I can recall correctly and the market has rallied ever since, therefore the average daily return is positive and there are likely to be more up days than down days whether the Vix spikes over its moving average or not. The Vix also made a secondary peak at just over 40 in March 2003 ( old VXO index) and has been falling ever since, so on average there have been more down days in the vix than there have been up days. Can you see the bias? What would be more interesting would be to analyse the relationship you have proposed from say 1995 to date so that you capture a period when the correlation of the VIX and SPX swung from postive until 2000 and then to negative from March 2003 to date.

Brett Steenbarger, Ph.D. said...

Thanks for your comment, Vegaman. You're absolutely correct: I am selective in the data I incorporate into historical analyses. That is because I am not seeking universal relationships that apply across all market conditions. (Indeed, I would argue that those relationships account for a very small proportion of variance in market outcomes). Rather, I am looking for local relationships that hold true under particular market conditions. Thus, the relationship between, say, the VIX and future price change might be quite different in a low-volatility, bullish trending market than in a high volatility, bearish market. By modeling only the most recent data and then by cutting the data sample in half and modeling each half to ensure uniformity, I try to identify those local relationships.

As a practical example, suppose I wanted to predict the outcome of the next Cleveland Cavs game. My past data sample for prediction would not go back many years, to before the LeBron James era. Rather, I would use the most recent data to model the likely next outcome.

This is the essential difference between the mechanical system developer and the non-linear historical modeler. The former is looking for universal relationships; the latter is looking for regimes that come and go. Clearly I'm in the latter camp.

Thanks for the opportunity to clarify an important but somewhat difficult issue.