A.I.M. Users Bulletin Board

[Conrad]

AIMster, on optimization!

By default the AI optimizer starts from 1 year ago to the present date, lets you plug in a stock or fund symbol, starting amount, then, under the genetic version, optimize for greatest portfolio value, greatest number of shares, greatest portfolio control value. (Choice of one, not all of the above!)

I gather you refer to the AI program of Mark Hing? Way back when Mark was developing his optimizer I discussed with him various means and methods to develop a subroutine for the Vortex Program. He gave me a lot of advice on how to optimize the investment using historical data. I believe strongly in the effect of this sort of optimization but it is only effective for stocks that exhibit, to some extend, a repetition of its +/- pattern and a program(or the investor) needs to identify that repetition (otherwise to keep optimizing is useless). In the end to create a optimization package for Vortex proved far to difficult for us and I abandoned the idea. Instead I stuck with manual optimization using an Excel version of Vortex and that work great but it is a lot of work.

I had the further thought of what if one started the optimizer from the date one initiates a program and always uses that as the same starting date? This would in effect allow you to tune your settings to optimized values as the program continues, rather than using the past to project into the future. Perhaps not the Grail, but it will be interesting to see if this adds value.

This would only be effective if the stock prices retain their pattern from the starting date. That almost never happens. Usually the pattern (if it exist at all) exhibits it self over relatively short periods. Keeping the starting date for the optimization fixed could create very wrong parameter settings that will last for a relatively long time, unless one uses an exponential decay method that reduces the effect of the past. If one does that then automatically one have an optimization technique that only considers the data of the recent past. . say for 1 year or less(selectable), if you recognize that then it is pointless to keep considering data from 5 years ago. Using a short data set makes the optimization faster anyway.

The scheme I wanted to use is an automatic package that would start with certain period of historical data and then use the optimized parameters to start the investment. Then as you go the optimization package would accept the stock price data as you enter the new data as you go, buying and selling as usual, but of course also entering data periodically for which you make no trades. For every set of data for a new period that is entered in the databank the optimization tool only used the latest data period. In a way you always use say a one year’s worth of trailing data for the optimization. You could do an optimization say every month or every week. With this type of optimization one could still use the exponential decay method for the data (weighing the present data more than the data from the past. . .this way you pick up on changing patterns more effectively.

It appears that AI does the optimization with the trailing data as I had in mind . . . possibly even suggested to me by Mark Hing

Do you know if he uses the exponential decay method for weighing the price data?. . .I never understood fully what his "genetic" method really meant.