Is it possible to create a system which will turn itself off, i.e. stop giving trade signals, when those signals are more likely to fail, without any intervention on your part or any attempt to use an external filter to detect market shifts? How would you even know if the feature was working, other than by decreased frequency of signals? Would something like this be a case of trying to "prove a negative" by concluding that if your system is giving you fewer signals, it means that it stopped giving you signals which would have failed. Anyone who's consciously tried to achieve this, please chime in with your results. I've been going through a period with a major decrease in trade signals, but the signals I've been getting still have a nice profit factor, so it's not as if the system has started to fail. Is it possible that this system has this feature built-in without me explicitly designing for it?
Unless you have code that does this, what you're asking is dependent on what your criteria for unfavorable conditions is, and if you know already, it can be coded. Seems obtuse and abstract. Everything you code is part of your system, and if you know what your conditions are you can backtest them to know if they improve your results. We have larger positions in NQ sometimes when two of our systems agree. 1 system trades 2 NQ, another 1 NQ, but when they agree we'd have 2 NQ and 1 NQ as normal and do 2 extra NQ. After backtesting that there was the better payoff of 0.01% more drawdown for 80+% higher returns. Not really a turn off. Backtesting any changes will let you know for sure. Without backtesting, though, this is not possible to answer. You might do that occassionally around news events such as FOMC announcements or GDP or unemployment reports, identify the dates of those and see if abstaining from action would have been prudent. Other than news events though, you should be able to code and analyze what makes you better off in terms of expectancy. The manual ommission of those dates can also help you improve your results by being more selective about market conditions. I think one I follow is never trade the friday before a 3 day weekend, especially if it's an early close. It just isn't worth it.
There is a rather simple trick which works for most (not all) strategies: You need a platform that is able to simulate trades not only in a backtest, but also in real time. Run your system and build a Moving Average of your equity curve. When the equity falls below the Moving Average, switch from trading to simulation, until the simulated equity curve crosses its Moving Average again from below. Then resume trading. This will, with some lag, eliminate unprofitable periods from your strategy. When you use a lowpass filter instead of an MA, you can minimize the lag. This method works because the equity curve normally reflects quite well the adaption of your system to the market.
My personal experience with such "switches" is that they can "whipsaw" you badly ... they get you out after a loss, then you miss the winners, and then you get back in again in time for more losses! ... my own conclusion has been that it's best simply to keep trading the strategy, and not try to advance pick the trades that will be winners versus losers. If you're comfortable the strategy is still working, that is ...
It depends on the strategy, or rather, on the "trendiness" of its equity curve. When the market situation does not change in a chaotic way, such a switch can often reduce losses. A typical example where some switch algorithm would work well is this strategy: http://www.elitetrader.com/vb/showthread.php?s=&threadid=242134
Yes, I was wondering more about "unintended consequences" of specific design choices in the system, which may or may not be covered by backtesting. In my case, I thought I had a system which gave about 1 signal every 2.1 days, but over the past 4 months, that signal frequency is way down to one every 2.9 days, but the profit factor is still more than sufficient to say it's a successful system, although not as high. Also, I was thinking about this question in the context of the "typical" system lifecycle which runs from solid backtesting results to either adequate or inadequate real-time results to unprofitable system (if real-time results were once adequate) and wondering if there are systems which never reach the "unprofitable" stage because they stop or slow down their signal frequency, but don't give "false positive" signals, which are really the primary way a system becomes unprofitable.
That's the thread I was thinking about when I posted this one. As I thought about it, I wondered if it was possible that there is some set of systems which don't need "switches" to detect market changes in an external manner because the inner workings of the systems' designs are themselves "switches" in some indirect way that eliminates the need for an external "switch". But that starts to get maybe too hypothetical because in some sense you are asking if a system designed differently from your own could have performed as well as your own did in market context "X" but then would that same different system fail in market context "Y" while yours didn't fail but did reduce its signal frequency in market context "Y". Then, would that fact say anything about your system's future performance in market context "Z". Or, more generally, could there be a system which doesn't "fail" per se, in the sense of ultimately returning your equity curve from whence it began, but rather stops giving you any signals, leaving your equity curve basically at its highwater mark indefinitely. I would think that if there were a way to consciously design a system like this, that a lot of people would be interested.
This is exactly the definition of a profitable system. ( Conversely you could say the system turns itself on when its signals are most likely to succeed.)
On a snapshot view, yes, this would define a profitable system. Considering the system lifecycle, though, would a system shutting itself off indefinitely be common? Seems like most of the systems you read about end up eventually in a state where they give signals, but the signals stop working and the equity curve goes back to zero net gain, with the user either riding the curve back down or ignoring the signals, depending on their willingness to stick with a now-failing system.