|
|
|
|
 |
Daily Speculations The Web Site of Victor Niederhoffer & Laurel Kenner Dedicated to the scientific method, free markets, deflating ballyhoo, creating value, and laughter; a forum for us to use our meager abilities to make the world of specinvestments a better place. Write to us at:
|
Philip J. McDonnell
4/26/05
Averaging Portfolios, by Dr Philip McDonnell
Thanks to Dr. Zussman for posting his links to academic studies on relative strength. The Miffre Rallis (MR) study uses the Jagadeesh et al methodology of averaging portfolios. MR average portfolios of strong and weak commodities for various lookback periods. They form portfolios by at least three successive levels of averaging:
1. Average of 3 highest quintiles or 3 lowest quintiles. (Note middle quintile overlaps).
2. Equal weighted average of portfolios for the last n time periods. Note this implicitly introduces a lag of n/2 into the results.
3. Summary results presented are yet a third level of averaging across time periods.
The final "momentum" portfolio is the difference of the winners and losers portfolios, yet a fourth level of averaging. At each level of averaging the standard deviation is implicitly reduced. It does not appear that the study adjusts properly for this and thus the t statistics are suspect for this reason.
MR also suffers from a multiple hypothesis problem, with at least 16 for the short term study. There appears to be no adjustment made so again the t-statistics cannot be believed. There is a similar problem for the long term study.
The study uses rising and falling interest rates to distinguish between the times to buy momentum and to sell momentum. Correlations are positive between returns to the momentum trials and interest rates, possibly indicating that the real correlation is between commodities and interest rates and not price momentum. The study doesn't rule this possibility out.
The Hsu Kuan study on TA carries the multiple hypothesis problem to a new extreme. They study nearly 40,000 hypotheses and not surprisingly find some that work. They attempt to test using various data snooping reality checks. Personally I don't find the tables of only the top strategies convincing. It would also be helpful if the authors had included the parameters of the top strategies so that others could test their results out of sample.
Note: I wrote this over the weekend but couldn't post it due to a lightning strike knocking out all of our computers and a hub. The lightning struck a fir tree directly across the street and ignited a fire.
(address is not clickable)