I just realize my ninja 7 xml template uses MAAverage which is not a built-in, so here it is. Note: import this one then reload the template.
Thanks for all the input & ideas. I'm looking for something that can be automated (for backtesting and scanner purposes) so I can't follow some of the more subjective approaches (double tops, etc...). My current implementation, which seems to be reasonably effective, computes a ratio of absolute value of movement of the indicator to absolute value of indicator's mean (using an EMA). This ratio gets very high if the indicator stays mean-stable, which by definition implies that the indicator isn't persistently staying above/below it's moving average (if it did, the MA would move). The ratio gets lower when the fundamental indicator seems to be trending to a new set point. Anecdotally this seems to give me a good solution, but I still need to backtest a trading strategy that incorporates it to see if it boosts performance. Attached are the code snippets if anyone has interest
CCI will do it RSI will do it Any oscaillator will do it You won't make money over time, however, unless you can predict ahead of time which lookback period to use. There is also a situation where your indicator can say that price has reverted even tho price didn't really move anywhere, thus potentially leaving you in massive drawdown and an indicator reading of "0". See this thread for more info: http://www.elitetrader.com/vb/showthread.php?s=&postid=3192028
Your diagram in that other thread is spot on, and is exactly what I'm trying to quantify. I'm pretty confident that within my basket of potential targets there are some that tend to chop up and down around a steady level, and others that do more the uphill climb like you show in the sketch. It's for exactly that reason that a standard oscillator can't distinguish between mean reversion caused by a move in the indicator vs. mean reversion caused by moves in the mean. I'm looking for other approaches to distinguish between these two cases, such as the above mentioned ratio of movement of the indicator vs. movement of the moving average. The LinReg suggestion also seems like an approach to this. Regress the size of displacement from the mean with the next bar's change of displacement from the mean. If there's a good reversion relationship, this should appear in a negative correlation between size of gap and movement of indicator (not change of gap). My interests are pretty specific in augmenting my particular models, but this seems like grounds for a useful generalized reversion probability indicator
FINALLY someone gets it Usually when I say indicators don't work, the replies are guru babble like "Just because you can't make them work doesn't mean they don't work" and then they try to sell me a course Really, the only way to do it is to: 1) have an infinite capital supply (so you can keep averaging down) 2) set hard stops after which you just say "forget it, I'm not going to win on this trade" Even Bollinger Bands (the first mean reversion indicator everyone tries) has that problem. Good luck man. I gave up on mean reversion stuff a while ago because, while I was better at it than I was at "trend following," the bottom line is that it comes down to being able to predict direction, which I cannot do.
I use a model that spits out the type of 'good' mean-reversion picture that you seek, but from your (understandably) opaque description I'm not sure how well it may apply. 'Fundamental data' ...such as an econometric time series? Periodicity? What does 'short-term' mean in your strategy? Are you able (willing) to upload a subset of the data, obviously disguising the actual nominal data value, in csv or xls format? I could quite easily run it through the model and graph the output.
How do you mean reversion guys deal with situations like this? http://www.elitetrader.com/vb/showthread.php?s=&postid=2075844&highlight=trend+followers#post2075844