Post by statstrader on 01/27/09 at 22:51:17
I would argue that most fat tails are caused by exogenous shocks or "jumps" in your model descriptions.
However, serial correlation would also likely have some impact on different time frames (ex: day vs. week variance ratios) as shocks tend to autocorrelate (GARCH) in serial fashion.
However, take away the fat tails, and the problem still exists.
I included a small hypothetical example here. System A has captured an underlying order in system B (the index it is applying rules on). The sign order in B is always - - + + XXXX, where the Xs are don't cares or don't trade because the signs can change in those variables. According to my backtests, system A is enormously profitable, since it never loses and always captures the relevant serial information that is embedded in B.
However, according to a T-test (w/assumed unequal variances), the system I devised is no better than chance. Therefore it is a failing system and does not reject the null that this system is no better than chance. This is the basic premise of the entire book as I understand it.
See any problems here? P.S. thanks
for keeping the threads alive. Wish more lurkers would opine.