Trading System Development 101 (Appendix B)

This year, I have been trying to get more organized by completing rough drafts into finished blog posts. Sometimes I don’t even understand what I have written because it has been so long, but I am presenting them anyway on the off chance someone out there can possibly benefit. In that vein, here are the last loose ends and notes regarding my mini-series Trading System Development 101 (concluded here).

—————————

From the last paragraph here, I could also look at what percentage of iterations are profitable when grouped by VIX cutoff value. I could then know how often a VIX filter would actually work and whether I get those desirable high plateau regions.

This post had a footnote where I indicated some further explanation could be useful. Looking back to that final full paragraph, imagine one set of trendlines might result in X, Y, and Z trades being taken. Were the chart to begin a couple bars later, imagine a different set of trendlines could result in A, B, and C trades being taken. Granted, multiple trendlines generated due to the allowable margin of error are better than zero or few trades (sample size too small). Both sets of trades are equally feasible, though, and should therefore be considered even though multiple open positions are not allowed in backtesting. Timing luck applies here to the trades themselves, as well as the trendlines with respect to where the chart begins and what bars will be available from which to construct trendlines.

The last loose thread I wish to tackle is from the final paragraph here: why is KD’s most common response to me “there is no right or wrong answer?”

This is an example of a standard response I get:

     > It could be hurting, hard to say for sure. Try aiming
     > for 100-200… a few times and see what happens. Or
     > even try 1000 or more. There are some who usually do
     > 10 or less, some that keep it under 100, and some that
     > always have thousands. So, there is no set answer to
     > this, because all can work (and all sometimes don’t).

This almost sounds like Yogi Berra wisdom!

My response is:

     > You say there’s no correct answer, but it may be an
     > empirical question. You could track the lifetime of
     > viable strategies (how long until they break). You
     > could then look at strategies with few and compare to
     > strategies with lots. Track how long until they break.
     > Compare the two groups to see which is longer.

He certainly could do this and I think it would be quite insightful.

However, recall his business model I detailed in the second paragraph here. Anything tested by others are strategies he doesn’t have to test himself. He will never know what everyone tests, but the more strategies tested by others, the more viable strategies will be passed to him. The more diversity in strategies tested by others, too, the more noncorrelation he can realize. The last thing he’d want to receive are strategies similar in one or more ways.

Discouragement of any kind is therefore not in his best interest. Whether it has few or many iterations, optimizes over this or that range, uses this time frame or that one, is mean-reverting or trend-following, etc., as long as it passes his criteria, it’s a strategy he will be very eager to check and/or implement for himself.