For mean-reversion traders, how do you guys arrive at your stop loss targets? Some multiple of the deviation from the mean, or just some empirical result that backtests while? Does it make sense to use time-based approaches? i.e., if the reversion hasn't happen in such and such time, just cut it -and- it has moved against you regardless of the result.
Hard to explain specifics, but my system dynamically moves the stop and calculates an estimated "edge per trade" left in the trade still. It's not a good system, but the losses are bounded and the pnl distribution skew is favorable. However, when I run the system in practice, 70% of the time I can tell when the trade is going to be a loser just on account of the excessive duration involved in the trade. This is why I am researching alternative stop loss mechanisms.
That would be a form of data-snooping. If your original hypothesis (mean-reversion) doesn't produce the results you wanted (the pair continues to divert from the mean), you should document it, and move on. Drilling-down and curve-fitting are BAD, but you already knew this. Right?
I did know this, but that's why I'm posting. The pursuit of edge-improvement. I mean at the tick level, I can see when something is not right by virtue of reading the tape. It's just that if it's obvious to me intraday, why wouldn't I want some generic, applicable methodology to identify the situation and cut the loser more quickly, before it balloons into a disaster?
The only time stop I use limits me out at breakeven ( if still down after X amount of time ). I find that usually my target will be hit or missed within 3 bars.
The price generation models tht you use will generally determine your paradigm. So you first make sure you have the stylized facts covered and then implement proper betting strategy. if you cover all possibilities of what the market does with different models, that's one way.
Ok, so say I have a histogram of profits and losses from simulation and real results. I break out each result into two columns and write down some facts, like time of execution, time held, etc. I do a kind of qualitative regression of sorts on the outcome and find that time-in-trade happens to be a significant "predictor" of loss. Now, there are two paths to take: 1) Use time itself as an input into cutting losses, or 2) Use time as an input into detecting that there's been a sort of regime change and specialize the strategy to deal with this Do you see any systemic problems or serious flaws with taking this kind of approach?