Ray Barros recently posted some worthwhile thoughts regarding the limitations of quantitative analyses of the markets. His basic point was that this bear market is different in size and duration compared with past bear markets. If we use data from the recent several decades to model the present market, we're apt to come to erroneous conclusions.
On the heels of posting about historical patterns involving price targets, I thought I'd explore Ray's idea with recent data.
The S&P 500 Index (SPY) traded down for the most recent five days and the last two sessions were weak, touching the S1 price target. Going back to 2000, when this has happened in the past, the next five days in SPY have had a bullish bias. Specifically, the index has averaged a gain of .36% (266 occasions up, 211 down). By comparison, the remainder of the sample has averaged a loss of -.24% (903 up, 908 down).
That might lead us to look for a bounce in SPY over the coming week.
If, however, we divide the historical sample into two portions: from 2000 through mid-2007 and after mid-2007, we see a significant difference. Prior to mid-2007, a down five-day period in SPY that ended with two days hitting their S1 downside target resulted in an average gain of .46% (208 up, 153 down) over the next five days. Since mid-2007, however, this pattern has led to an average gain of only .03% (58 up, 58 down).
One implication of this finding is that a good part of what sustained the bull market was the tendency of short-term periods of weakness to be followed by short-term strength. A good part of what is sustaining the bear market is the loss of this tendency.
When we look at the reverse pattern since 2000--an up five-day period in which the last two trading days fail to touch their S1 downside levels--we see that the average change over the next five days has been -.30% (277 up, 306 down). Since mid-2007, this same pattern has led to an average five-day loss of -1.47% (29 up, 58 down)!
Notice that the lookback period of 2000 to the present does indeed include bear as well as bull periods, as the market was lower from 2000 through early 2003. Nonetheless, it's clear that the market patterns since mid-2007 have shifted from those that preceded. Weekly weakness has not been followed by a strong bullish edge and weekly strength has been followed by substantial weakness.
While this doesn't invalidate the use of historical data to frame trading hypotheses, it does underscore the importance of examining the recent data relative to the past. When we see a historical trading pattern failing to hold up in the most recent portion of the data set, it's a good sign that market regimes have shifted. In that situation, the best guide to the future is the most recent past, not a different, more distant period of market history.
.
Subscribe to:
Post Comments (Atom)
Posts
2 comments:
Brett ~
This is an insightful post about analysis of all kinds, not just that founded on quantitative measures.
I think you would agree that there is a near overwhelming need on the part of all market analysts, whether technicians, fundamentalists or quants, etc, to have their insights verified.
This need for self-verification frequently drives people into trades they would not have been in under purely objective circumstances. Some wise man expressed this as, "The need to be right rather than make money."
The need for verification enters market analysis at the outset, at the model design level, not at the back end where the money's deployed. Many models ~ technical, fundamental or quant ~ are designed to affirm previously held biases.
Writing as a quant, I can affirm the truth of this sort of bias entering mathematical processes at trading desks worldwide. That's right, mathematics is not a strictly objective practice. The next time any quant tells you it is, snicker and take his money in the countertrade.
Ray Barros makes an interesting and valid observation in his post, and you expand eloquently upon it in yours.
In response I say that proper model design ~ design intended to FALSIFY one's hypothesis, not to verify it ~ would make use of valid statistical tools compensating for varying period lengths, varying volatility and volume, and perhaps even the change in composition of market proxies.
Anyone can make a thing complicated, so the true puzzle for a quant is maintaining model simplicity while maintaining robustness in the face of temptation to feel like a real math stud (each one of us gets his pleasure where he can).
For a model to demonstrate robustness, it must first survive falsification and then return repeatable results under multiple independent trials. This ability would satisfy Ray Barros's valid objection.
Unfortunately (fortunately?) the emotional need for self-verification prevents the majority of market participants from obliging their ideas to undergo such rigorous testing.
Adam.
Great, great perspective, Adam; avoiding confirmation bias is a major challenge for any discretionary trader. Thanks for the note--
Brett
Post a Comment