Sunday, September 22, 2013

lessons of ai: aiSpy 6 months later ...

last spring, i unveiled a study called aiSpy that i had created (with a good deal of effort mind you) that implemented a form of artificial intelligence called a feed-forward network. this feed-forward network was static, that is it didn't modify itself but was the end result of an extensive genetic training program. ffn's are used to do complex pattern recognition for speech-to-text and handwriting recognition and many other human-interface applications you may be familiar with in your cell phone.  if there were patterns to the complicated dance of the market then they ought to have ended up encoded in the ffn, which at the time of release was outperforming buy and hold over 2, 5 and 10 year periods.

so how did aiSpy do in the real world? here's a picture of the performance over the last 6 months, post-training:
last 6 month performance of aiSpy.

the p&l for aiSpy, trading $10,000 per trade, is circled, 497.08, and contrasted to the growth of $10,000 bought 6 months ago and held (bnh, green number in square)  and the growth of $10,000 invested in 6 increments (dca, blue number in square.) bnh is the winner with a $967 gain, dca came in 2nd with a $545 gain and aiSpy third with a $497 gain. i should note that the aiSpy p&l does not include the impact of commissions, which at the standard thinkorswim rate amounts to $10 per round-trip, or $120 over the last 6 months for 12 twelve trades. the aiSpy result is an even more distant 3rd. while this is a lack-luster result, at least it is positive. what went wrong? i think i know and this knowledge is helping me improve my trading and i will share this in a series of posts called lessons of ai.




No comments:

Post a Comment