Mean Reverting Strategies

Recently there have been some discussions on one of the quant groups, where someone indicated “mean reverting strategies no longer work”.    I don’t agree with this at all.    However, there has been a lot of risk aversion and the market movements of the last few months (well since Feb or March) have been quite different then those preceding.    The market of the last few months has been ideal for market making given the tight bands and much less favorable for strategies that rely on trends or big swings.

One of my strategies is quite simple and depends on mean-reverting behavior across a large portfolio.   When MR behavior in the market is “cranked back” it makes sense to filter the portfolio, increasing the chance of successful trades and reducing possible losses from assets without the desired footprint.

I decided to do some tests on the biggest winners and losers by asset to confirm my belief that wining assets would exhibit strong negative autocorrelation and losing assets would likely exhibit positive or statistically not-significant levels of autocorrelation (+/-).  Indeed one of the biggest losing assets had the following profile:

The above asset spends most of the time in positive auto-corr territory and does not tend to mean-revert over the window of interest.    The asset below spend most of its time in negative auto-corr territory and hence is very strongly mean-reverting:

Of course measures like this have lag associated with them, but if chosen carefully can be used to effectively filter assets on the run.   Here is the test code in R:



Filed under strategies

6 responses to “Mean Reverting Strategies

  1. Great post. Out of curiosity, what time frame did you use when testing for auto-correlation ?


    • tr8dr

      I inlined a screenshot of the code in the post now. Not sure which tag is best to use for code in wordpress. The “

      " tag text is very large in this template.
      I saw your work on volatility estimation.  Have not read in detail yet.  How predictive did you find GARCH to be?   I've mostly focused on intraday prediction in the past so don't have a read on how well GARCH does for daily (it is not effective for intra-day but papers indicate good for daily).
      Which variant of the model worked best for you?   Also I noticed that you are looking at vol of vol.  Come across any models that make use of vol of vol for vol forecasting?
      • Thanks for the reply. I found the GARCH to be fairly good. I used the (1,1) model for the post, and generally I had better results keeping the model at a low order.

        I personally never tried it on the intraday time frame yet so I can’t really tell, but like you said papers were suggesting good results and from what I tried so far, it worked fairly well and in-line with my expectations.

        Regarding vol of vol, I am still digging into it. There is such a high correlation with vol that most models I tried are not useful as is. Tony Cooper suggested to get rid of the correlation effects, he was mentioning promising results on FX. I am still in the research process but haven’t had much luck yet to be honest.


  2. Autocorrelation is also something I’ve been looking for. Significant autocorrelation gives an indication that the time series should be predictable. From the time series I’ve seen so far (pair and single stock) most exhibit very low ac (<0.2), SPY is one of them . Funny thing is that some people believe that some indicator could predict future of a random walk :).

    • tr8dr

      What are you using to measure the autocorr? I am not using the autocorr to fit into an arma model rather just using historical autocorr as a filter on which assets I will use for a completely different MR model.

      You should be able to see auto-corr for an asset that has short-term linear-style MR.

      The assets you are looking at may exhibit non-AR like MR. Another test one can consider is the variance ratio test …

      • sjev

        I was not quite happy with the results from fitted arma models, so my current use of acf is similar to yours, as a pair selection filter.
        Regarding the MR behavior in general, I must say that no matter what strategy I try (etf universe), they all have diminishing returns since 2008. Extrapolating the rate of performance decrease I think most of them will become unprofitable in a year. However, there have been enough volatility in the past months and I’m sure there are many opportunities to take advantage of.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s