Recently, blogger Barry Ritholtz kicked off a discussion of the historical patterns analyzed on this blog. Most of the analyses extend from March, 2003 to the present, which tracks the recent low volatility bull market. Shouldn't we analyze over longer market periods? Would the historical patterns remain the same, or do they change over time?
As I indicated in my responses to these questions, the answer to this question differentiates those who are searching for universal relationships in the market (mechanical systems that will work in any market condition) from those who continuously model and remodel markets to find local regimes (relationships that wax and wane over time).
Here's a practical example: Let's say that I am trying to predict a person's behavior in response to a piece of news. I could take a large past sample of behaviors following receiving news and use this to make my prediction. That would be a linear approach, looking for a universal relationship between news and behavior.
A different approach might say: The person's behavior in response to news will depend on their moods, which come and go. So I will have one prediction if the person is in a good mood, another prediction if mood is neutral, and a different prediction when the person is in a bad mood. My samples for analysis will consist of only good mood occasions (for my first model), neutral mood occasions (for the second model) and bad mood occasions (for my third model). This would be a non-linear approach, finding different relationships between news and behavior as a function of mood.
Consider that volatility and trending are variables that define market mood. If that is the case, the non-linear historical modeler would argue, relationships between indicators (such as the VIX) and future price change should be specific to the mood (market conditions) at the time. There is no strong, universal predictive relationship across all moods (market conditions).
Because of this, no trading edges last forever in the view of the non-linear modeler.
Let's take the relative VIX as an example. In my previous post, I found that, when the VIX exceeds its 20 day moving average by 15% or more, there is a bullish edge going forward. That analysis covered March, 2003 to the present. But let's extend the analysis from January, 1998 to the present (N = 2124 trading days).
During that time, we have 159 occasions in which the VIX exceeds its 20-day moving average by 15% or more. If we divide the sample in half based solely on time, a pattern emerges. Between January, 1998 and March, 2001 (N = 80), after such a VIX spike the S&P 500 Index (SPY) was up by an average of .62% two days later (51 up, 29 down). This outperformed the overall sample, which shows a two-day average gain of .04% (1111 up, 1013 down).
From March, 2001 to the present, following a VIX spike (N = 79), the market was up by an average of only .05% (43 up, 36 down) two days later. This does not outperform the overall market.
In other words, the relative VIX gave us a bullish edge between 1998 and early 2001, but not afterward. From the earlier post, we saw that the VIX spike gave a bullish edge from 2003 to the present. As you might imagine, the same VIX spike provided no edge--and actually led to a slightly bearish outcome--during 2001 and 2002 (-.07%; 25 up, 25 down).
We could come to two conclusions about the VIX from this little investigation: 1) it is a poor indicator, since it does not yield consistent predictions over time; or 2) it is a good indicator, as long as we utilize it when current market conditions suggest it will be effective. Clearly, this site takes the latter approach. It's not a foolproof method--market cycles can possibly change without our being aware of it at the time--but until we find the perfect, universal indicator, it might just be the best we can do in the face of complexity and uncertainty.