Mining for Trading Strategies (Part 2)
Posted by Mark on February 15, 2019 at 07:11 | Last modified: June 17, 2020 13:48Today I am going to mine for more trading strategies using the same procedure presented in Part 1.
Today’s strategies will be short CL. See Mining 2 for specific simulation settings.
Seven out of the top 20 (IS) strategies passed the Randomized OOS test (see fourth paragraph here). None of the seven had a PNLDD > 1 in the 4-year incubation period, or a PF > 1.24.
Six of the top 30 had an average Monte Carlo drawdown (DD) less than backtested DD. Only two of the six passed Randomized OOS. I did not run the others through incubation.
Not happy with this, I ran another simulation (see Mining 3) for short CL strategies the very next day.
Nine out of the top 28 (IS) strategies passed Randomized OOS. With PNLDD > 2 and PF > 1.30, two of these nine passed incubation. Five of these nine had an average Monte Carlo DD < backtested DD, but none of those passed incubation. Thirteen of the top 28 strategies had an average Monte Carlo DD < backtested DD, but none of these passed incubation.>
Still not impressed, I changed the exit criteria slightly for my next short CL simulation (see Mining 4).
Twenty-two of the top 31 (IS) strategies passed Randomized OOS. One of these 22 passed incubation (PNLDD 2.44, PF 1.57). Five of these 22 had an average Monte Carlo DD < backtested DD, but none of the five passed incubation.>
Running this simulation generated 22 strategies that passed Randomized OOS but only one that passed incubation (and none that passed those + Monte Carlo DD). I questioned the utility of looking at so many Randomized OOS graphs since such a small percentage pass but on second thought, viewing more is necessary for exactly that reason.>
I have seen strategies pass Randomized OOS once but fail to repeat. With Monte Carlo DD, I run the simulation three times for confirmation. I will do the same for Randomized OOS going forward. I also thought about requiring both OOS and IS equity curves to be all above zero for Randomized OOS. The hurdle is high enough already, though, so I will hold off on the latter and just focus on requiring multiple passing Randomized OOS results to confirm.