Filtering Trade Signals

[blinkybill]

I have been playing with the TMA system and applying various filters such as correlation and unit quantities. Obviously the system can be improved by tightening or loosening the various filters to allow more trades through. I was wondering how other people normally handle this tradeoff? Do they simply go for that which maximises their objective function be that return or MAR or are there other factors at play? In my case the best result (MAR) is achieved by the tightest filter which knocks out about 30% of all trades. But then a test with a filter knocking out about 10-15% of trades isnt soo much worse..I wonder whether knocking out so many trades in a LTTF system doesnt mean I miss so many wonderful trades....the results seem to suggest not but then maybe..I dont know.

Second part of my question is do many people then go to the next step and try to rank the quality of the signals that are filtered out versus those that have already been accepted? e.g with a correlation filter do we compare the filtered trade to an existing correlated trade that has been on for say 50% of its "normal" trade life and chose to remain or replace as the case may be....is there any evidence to suggest this kind of approach is useful.

Thanks Jim

[sluggo]

I find it comforting when a filtering rule affects 35% to 65% of the trades (thus it does NOT affect 65% to 35% of the trades). This quiets my nervous fears of overfitting, and helps me believe the rule wasn't accidentally tuned to merely polish up the top 5% of big winners (or eliminate the bottom 5% of horrible losers) of the in-sample period.

Sometimes people put in filters "just because It Is The Right Thing To Do" and disregard their effect on the objective function. One example that jumps to mind is Volume Filters; if recent volume is too low, don't take the entry signal, regardless of how profitable such trades may or may not have been in backtesting. Other times, filters are added because they control something that isn't necessarily "THE objective function", but is important nevertheless. Such as the Correlated Units filter, whose job is to reduce Catastrophe Risk rather than to increase the R-Squared.

[7432]

for me the tradeoff between risk and return with total and correlated units comes down to being in the game for the long term.
I like to test a system against shock moves when everything becomes correlated and tighten the positions taken so that I think I could have handled the big drawdown.
it might mean giving up quite a bit of return on a back test of 5 or 10 years, but lowering the drawdown means I'll be around the next month.
gets scary when cotton, ibex35 and hogs all become correlated, but that seems to be the case recently.
watch out for those highly negative correlated units too.

[LeapFrog]

The second part of your question reminds me of the "parking lot" discussion that can be found elsewhere on the forum. Forum member and hedge fund manager Dean Hoffman seems to have developed a system doing just as you say, although the details are black boxed.

Still, his fund performance would suggest that things are working out well for him and that his approach therefore might have some merit.