Hi Dave
You are right, an ema as a threshold for transactions is an interesting wrinkle, but one which, near as I can guess, somehow just might violate the underlying principle of AIM itself. Think of it this way, AIM is really automatic when you just use its moving attractor basin as the signal, but when you throw in something that says "I don't care if you want to buy <sell>, you're not below <above> my threshold." Kindof like meddling with Moses. However, where cash flow (read: trading performance) comes in, I sortof throw out the rules and see what really works. That's evolutionary programming.
If I recall rightly, I did a variety of trials under EP: some mutual funds, some stocks, and always my favorite stock ISIP. I used several years of data. What I did not do was to use "walk forward" training/testing. I just settled on long batches of train followed by test by dividing available data into two groups. Walk forward might have given even better results because you evolve (read: learn) more often. That will be included in my next round of experiments.
What I did was let ep fiddle with a variety of things, include buy/sell thresholds, and initial cash ratio. Didn't fiddle with initial funds: that was given so that all trials started with the same funds, but varied what was done initially with those funds. All the other variables illustrated in the paper were subject to EP under an initialization screen where you "checked" who could be mutated. I always used the same settings.
Hope that clears up any confusion. Cheers, Jack |
