Crude Oil Strategy Mining Study (Part 4)
Posted by Mark on August 28, 2020 at 07:49 | Last modified: July 16, 2020 07:44Last time, I presented results for the first factor: R(ules). Today I continue presenting results from my latest strategy mining study on crude oil.
The second factor, Q(uality), has a significant main effect on performance (see hypothesis[2]). Not only are best strategies an improvement over worst, the best strategies are [marginally] profitable. Top strategies (n = 816) average PNL +$87, PNLDD 0.34, Avg Trade $1.38, and PF 1.05. As mentioned in the third-to-last paragraph here, none of these numbers approach what might be regarded as viable for any OOS results. They sure would be damning if they landed on the side of loss, though.
D(irection), the third factor, has a significant effect on performance: Avg Trade -$17 (-$57) and PF 1.01 (0.94) for short (long) trades. This should not be surprising since CL fell from $84.60 to $54.89 during the incubation period creating a downside bias (hypothesis[3]). What should be surprising is that as with PNLDD and Avg Trade in Part 3 (second table), Avg Trade is negative for both long and short trades while PF < 1 only for the latter.>
A glance at the variable distributions helps me to better understand this apparent sign inconsistency. Avg Trade has an approximately Normal distribution with sub-breakeven (negative) mean. PF has a skewed distribution with a right tail out to 2.37 and left tail down to 0.22. In other words, the right tail goes 1.37 units above breakeven (1.0) while the left tail only goes 0.78 units below breakeven (1.0). This should produce some upward pressure on average PF to exceed 1.0.>
The fourth factor, P(eriod), does not have a significant effect on performance. I find it peculiar that the 2007-2011 training period generates significantly more trades than 2011-2015, but I’m not sure why and I don’t think it really matters. >
A significant 2-way interaction effect is seen between R and Q. Performance improves slightly in looking from two to four rules across the worst strategies whereas performance declines much more when comparing two to four rules across the best strategies. This interaction is the first graph shown here.
A significant 2-way interaction effect is seen between Q and D. Performance improvement is much greater for top vs. bottom strategies on the long side whereas performance improvement is marginal for top vs. bottom strategies on the short side (like the second graph in the link provided just above).
A 3-way interaction effect is seen between R, Q, and D. This is significant for Net PNL, PNLDD, and Avg Trade and marginally significant (p = 0.057) for PF. I’m not going to try and explain this interaction nor am I going to undertake a 4-way ANOVA by hand to screen for a 4-way interaction that I probably wouldn’t understand either.
I will continue the discussion next time.
Categories: System Development | Comments (0) | Permalink