Walk-Forward "cluster analysis"

Discussion in 'Strategy Development' started by BrooksRimes, Sep 4, 2010.

  1. This is an interesting and useful feature that this software contains. I've never heard of it in any other package and wonder if anyone else has heard of anything similar?


    P.S. This product has been purchased by Tradestation and will eventually be incorporated into that product.

    "We call this comprehensive testing of multiple walk-forward analyses a cluster analysis."

  2. For 99.99....% of strategies this will produce curve-fitting. Take a look at the other thread


    It will only work for strategies that cannot be optimized in the first place, thus it is not needed.

    IMO, this is what people tried in the 1980s. Now, if any broker wants to offer this optimizer/curve-fitter to their audience, the only reason for doing that would be to take the opposite side of fitted systems. Since I do not think TS does that, they are simply behind in their understanding of trading system technology.
  3. jalsck


    Interesting coincidence, I was just talking about the confusion that the misuse of the term curve-fitting creates on the other thread...

    The only way to avoid curve fitting is to trade at random.


  4. I read the other thread and I side with James.

    There is not a problem with curve fitting, there is only a problem with over curve fitting.

    My question wasn't on the merits of optimization. It was asking if others have found any other products which do a cluster-type walk forward analysis besides the Grail Systems product ...

    Would still like to know this?


  5. Brooks,

    This is one product I looked at briefly, but decided to purchase Grail.


    They also offer cluster analysis. I believe doing a stress test of the parameters which we spoke of in another thread and this cluster analysis, will help put the odds in our favor.
  6. Bear,

    Thanks. Looks interesting - many features. I'm not as keen on the leasing.

    Can I ask what your results have been with cluster analysis?

    Have you ever found a case where your single WFA failed but elsewhere in the block of 30 there was a profitable group?

    How long does an average cluster take to run?


  7. Most of my systems have 4 or less parameters, so cluster analysis usually completes in 3-5 hours. Grail on TS runs much slower than grail on multi-charts which is multithreaded.

    Yes I believe its ok for a single or even a few WFA to fail. Very rarely do I have all successful WFA in a cluster.

    The fitness functions are also quite interesting. As you know picking the right parameters is somewhat of an art than exact science.