Monday, January 18, 2016

aispy follow-up

one of my reader's asked me to follow-up on aispy, some 2 years after the release. so here's what this looks like today:

there's a $1400 loss vs. a $238 gain on a buy 'n hold. what went wrong? ai spy is a 'tuned' system. that is, it searched for settings that produced good results in a backtest. so if the forward test era was something like the backtest era we would have seen positive results. it is unreasonable to think that the market would maintain a correlation to any particular time period, thus, a drift in results is to be expected. i had no delusions that the market would stay in-tune with the backtest era. however, it's not unreasonable to think that maybe a small period of time following the backtest era might be positive and to be sure, if you scan back in the p&l subgraph there is some green on the screen. not much, though. also, since it is compute intensive to tune aispy (it takes a weekend of computing just for one ticker!) i have pretty much let this fall to the wayside.

the learning experience i had from programming an ai system was invaluable. what ai does is to cast your 'problem' into a vast solution space and then hunt for solutions that are acceptable or just better than what was previously achievable by pure reasoning alone. if these 'better' solutions abound in your solution space then the ai process will find them quickly. if not, then not. however, i don't think the ai 'knows' anything. that would require human beings to program the network weightings to somehow gravitate to special classes of solutions but then would those solutions be better than what could be achieved by pure reasoning?

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.