Hi guys, great forum with a lot of interesting topics!
It's been very interesting to read your thoughts on the "target exit topic", something I've had a very painful experience with. That is, in relative terms. We run a fairly large CTA in the Nordic region which up until last year showed relatively strong numbers compared to many of our competitors, a fact that slowely allowed our sales people to gain traction with institutional clients. However, early 2010 we introduced a new set of models in our portfolio, of which about 40-50% carried some form of target related exit. We felt pretty confident that they wouldn't be activated at the same time because they were so distant in time-frame and profit level. However, that was before the mother of all bulls showed up in bonds, effectively triggering profit taking in most of the systems...before the last leg of the rally began. Hence, we were running at half speed when most of our competitors were running at full throttle. We left about $100 million of our clients money on the table, basically giving up (and then some) the slight edge we had created over competition, in two months! For those of you that haven't really felt, or bothered about, relative pain I can tell you that it hurts almost as much as absolute pain
I would imagine that most people don't really pay attention to the above issues when simulating historic performance of a system. However, in the real world they may actually be of greater importance than the Sharpe.
A second issue I have with target exits (that is in a trendfollowing context, not when it comes to mean reverting stuff) is the fact that we introduce one or more degrees of freedom, everything else being equal. This obviously means that we should be able to extract higher Sharpe ratios in a backtest (we did, but actually diversification was our key motivator, anyway). But this doesn't necessarily mean that they will repeat in real life, or actually it's very likely that they will not. Hence, a backtest doesn't really tell us if the system has merit or not, if we don't adjust for the increase in datamining (sorry if I'm stating the obvious).
Lastly, which again has to do with datamining, is the extreme sensitivity to certain settings. Let's say that we use ATR to take profit and the level we use is 3.00. For obvious reasons a system like the one above will perform vastly different if one or more trades reach 2.99 ATRs versus 3.01 ATRs. Hence, a very slight change may render our system more or less useless, or at least the possibility exists. Compare this with the much less curve-fit moving average system. In that case it doesn't matter much if we pick 299, 300 or 301 days as the lookback (under most circumstances).
Just my 2 cents...