I am interested in learning more about how to optimize trading systems that are somewhat discretionary in nature, i.e. not easily programmed. Say for example I have an excel sheet which lists a number of variables (e.g. 12) and each time I enter a trade I record the value of each variable, assuming then at some point I develop a reasonable database. What techniques/methods can I use to determine which variables are the most important and which combinations of variables produce the best results? I guess I'm looking more for optimization of a database, rather than the traditional optimization/backtesting of an arithmetic/boolean rule set.
If you are optimizing on trade profit some sort of factor analysis should work (i.e. is there a relationship between Variables 1-12 and Profit). With 12 independent variables you would need a pretty big sample size, though.
Here are a couple of links for Multivariate Statistics in case anyone else is interested in looking further. The second link contains an interesting list of statistical software ranging from Freeware, to Shareware, to Retail. If anyone has gone this route I would be interested in some recommended reading and/or recommended software. Thanks.
I just googled optimizing rules and stumbled across this post. What you're looking for is statistical classification. Basically most trading systems are rules based but actually that's not how the human brain thinks. Take a look at Support Vector Machines (a form of statiscal based artificial intelligence) as this can do what you ask.
Take N columns of variables in Excel, they could be anything, prices, % increase in the last 5 days, % of price above an MA, Indicator values, recent high, recent low, previous day pivot point etc. etc.
Now take a final column with +1,0,-1 signals for Buy, hold, Sell. These you code in yourself and you use perfect hindsight to do it.
Finally run the whole lot through a support vector machine, with the N columns as training data and final buy/sell/hold column as output of the classifier. The SVM will best "learn" what is significant to the output. Hopefully if you did it right when you enter new data to the SVM (same N columns, minus the buy sell hold one) it should output decent B/S/H signals. Well thats the theory
I am interested in developing an algorithm along these lines. if you want to chat more feel free to PM me.