I've read that trend-following strategies perform well in high-volatility regime... and I've also read opposite statement saying that trend-following strategies perform well in low-volatility regime... But I seem to find it hard visualize intuitively either case. It seems to me "volatility" is not a good dividing line for good/bad performance of trend-following strategies. I am thinking of "auto-correlation"... but not so sure. If "auto-correlation" is a good dividing line for the good/bad regimes, then is there a way to "forecast" auto-correlation? How about mean-reverting strategies?
Trend following relies on directional moves, volatility can be high or low, you can still make money from a directional move... if volatility is low it will just take more time. Big moves are often preceded by contraction followed by expansion in volatility, this works across all time frames. Mean reversion depend on prices reverting back to the mean: you are buying low, selling high relative to your reference point. I've generally found mean reversion works better in low volatility environment and trend following works better in an expanding volatility environment but it depends what kind of systems and parameters you are using.
Adding volatility filters to improve the performance of trend following mechanical trading systems, is an old and time-honored practice. One system seller says I've come up with a trading filter that indicates when a commodity is becoming abnormally volatile. When the filter determines that activity is within a normal range, trading can take place with any suitable strategy. But when abnormal volatility is detected, on-going trades are exited and new trades are not entered in that commodity. I've tested this on a dozen or so strategies I monitor, and without exception the filter significantly reduces draw-down, and increases the ratio of profit to draw-down. Notice that they don't say whether "abnormal" volatility is volatility that is (A) too low; (B) too high; or (C) either too low or too high. So that's something for you to test. I'm sure you can think of two or three ways to quantify "volatility" and four or five ways to define "too high" and/or "too low". Program them up, attach to some of your favorite trend following mechanical trading systems, and try them out. You may get results similar to the claims quoted above. Or you may not. The only way to find out is to perform the tests and study the results. Assuming you do find one or more volatility filters which improve the performance of trend following systems, you certainly could use the filter to define "regimes". When the filter says "OK to trend trade", call that the "Trend Following Regime". When the filter says "not OK to trend trade", call that the "Mean Reversion Regime". Voila. You could do the same thing with your favorite mean reversion mechanical trading systems. Attach some volatility filters, run some tests, and identify the filter(s) that give the greatest performance boost. Call this your regime indicator. Voila number two. If you are absolutely dying to know the author of the quote above, operate the googles. They won't let you down.
There is a ton of academic info on this matter as it is widely accepted that volatility displays significant autocorrelation across all products/time-frames. Understanding regime shift means understanding the concepts behind volatility analysis and forecasting. There is some great practical info out there should you choose to take a thorough look. I'd venture a guess that any system that shows promise is actually capitalizing on volatility... why? Vola is an MR process that displays conditional heteroscedasticity. http://en.wikipedia.org/wiki/Heteroscedasticity Enjoy. Mike
Regarding the OP's original question: I discourage generalizations. However, generally speaking, you will find most experienced traders will fade highly volatile multi-sigma single day or event-driven moves and trend-trade the subtle grinding markets.
That's kind of an ignorant statement buddy... Ever hear of ARCH models? How about Heston models? The OP directly asked about forecasting auto-correlation. I gave him a few clues on where to start. Normally I prefer not to just give away answers and let people do their own homework. Jeez, its posters like you who remind me of why I rarely post on ET nowadays.
Hey, you're my buddy! I like having buddies... Anyway, do you really think OP is somehow going to make money by reading the Wikipedia page on heteroscedasticity -- the link you gave him...
No I don't think that the OP is going to make money reading about the H word, and I don't really care about that... however, I do know the OP has a history of asking questions and expecting the answers to be laid out for him. I'm sure that if Mizhael starts thinking and does some research about conditional variance, he might stumble across GARCH and Engle. Then he might actually read a paper or two that outline how to do some silly (and useless) volatilty forecasts. From there, he can come up with some different (and possibly practical)models that can produce decent conditional variances... I'll leave how this applies to systemic trading alone as I think most here do not have systems that can actually benefit from this type of approach.
Mizhael, order up a copy of "Modeling Financial Time Series with S-Plus" by Zivot and Wang. I have the second edition, there might be a later one. Surf the Wilmott forum threads, because quite frankly you are going to get lots of misleading and ill-informed discussion on financial modeling on this website. Personally, I am better at the trading aspect than the quantitative minutae and would be the first to admit it. My previous post was as lucid and consise as I could make it. Just trying to save you some time and frustration.