Trading System Development 101 (Part 8)

Last time, I introduced a data-mining approach to trading system development.

To summarize, here are the three steps to developing strategies the data-mining way:

1. Select: market(s) and test dates, entry signals, exit criteria, and fitness function
2. Run simulation to create strategies
3. View resultant strategies, stress test, forward simulate, create portfolios, check correlations, print tradeable code

Although this approach to strategy development does not require me to provide strategies, I am already anticipating an organizational nightmare. The simulations take time (proportional to complexity) and I don’t want any duplicates.

I need to come up with a system to label and track simulations. Each simulation will have entry/exit signals, profit targets, stop-loss, additional exit criteria, designated markets, direction, fitness function, number of trading rules, etc. Many of these selections are mutually exclusive (ME) and will require separate simulations. For example, different fitness functions are going to result in different strategies. I hope number of rules is ME. If not and selecting X rules means < X rather than X and only X, then I will have to give this one more thought. Long versus short is ME and will require a separate simulation.

The software also has other features that will give rise to additional simulations in need of organization. Minimum number of OS trades and/or percentage of total data allocated to OS can vary and give rise to different strategies. The software allows for intermarket signals, which at this point I have no idea how to categorize or test. I can say the same for ensemble strategies, which take positions only when designated combinations of other strategies have done the same.

Although the software creates strategies automatically, the rest has to be done manually. I can’t enter fitness criteria, which means I will have to sort on that based on pre-determined critical values. I will then have to run and eye the stress tests independently. Each stress test will probably give rise to an accept/reject decision. Any reject decision may be reason to move on to the next one. I’ll know more as I get into actual work with the software. In either case, I may want to document what stress tests were done and how they fared: more aspects of this grand organizational feat.