<html>
  <head>
    <meta content="text/html; charset=ISO-8859-1"
      http-equiv="Content-Type">
  </head>
  <body bgcolor="#FFFFFF" text="#000000">
    Hi Pierre<br>
    <br>
    Don't want to sound unhelpful but I'm afraid my employer would't
    want me to go further.   <br>
    <br>
    Have a look at Chapter 5 of my PhD Thesis
    <a class="moz-txt-link-freetext" href="http://www.xesoftware.com.au/ThesisAsPassed.pdf">http://www.xesoftware.com.au/ThesisAsPassed.pdf</a><br>
    <br>
    It is entitled <i>Experiments in Classification</i> and takes you
    through R and the various techniques.  True its in a rather
    different domain (speech recognition errors) but the problem is just
    the same: one of classification.<br>
    <br>
    <div class="moz-signature">Stephen Choularton Ph.D., FIoD<br>
      <br>
      9999 2226<br>
      0413 545 182<br>
      <br>
    </div>
    <br>
    On 24/08/2011 7:00 PM, Gentil Homme wrote:
    <blockquote
cite="mid:CAM9XErB3JXSK4TTCQpDP+iQy3jEWXVJfLawbyJQ1ezT_Q3zC=Q@mail.gmail.com"
      type="cite">Thanks Stephen,<br>
      <br>
      <br>
      I'm currently first attempting to define one architecture (inputs
      to NN, which I believe need some preprocessing/denoising, ....)<br>
      I think that the process to train and test NN performance must be
      well organised, and the validation steps are much critical ...<br>
      <br>
      I found also the neuralnet package that seems interesting for NN
      training.<br>
      <br>
      I'll look after svms in a second step.<br>
      <br>
      Which kind of data did you use in your nn tests ?<br>
      <br>
      <br>
      Best Rgds,<br>
      <br>
      Pierre<br>
      <br>
      <br>
      <br>
      <div class="gmail_quote">2011/8/24 Stephen Choularton <span
          dir="ltr"><<a moz-do-not-send="true"
            href="mailto:stephen@organicfoodmarkets.com.au">stephen@organicfoodmarkets.com.au</a>></span><br>
        <blockquote class="gmail_quote" style="margin:0 0 0
          .8ex;border-left:1px #ccc solid;padding-left:1ex;">
          <div bgcolor="#FFFFFF" text="#000000"> I use nnet and it
            appears to work reasonably 'out of the box' as an algorithm
            but you get better performance if you play with the
            parameters to tune it.<br>
            <br>
            svm appears to work best when you select Gaussian so I'm not
            sure about your assumption.<br>
            <br>
            All these algorithms have similar R calls so its not to
            difficult to try lots of them out.  Try then out of the box
            first up.<br>
            <br>
            good luck.<br>
            <br>
            <div>
              <div class="im">Stephen Choularton Ph.D., FIoD<br>
                <br>
              </div>
              9999 2226<br>
              0413 545 182<br>
              <br>
            </div>
            <div>
              <div class="h5"> <br>
                On 23/08/2011 4:51 PM, Gentil Homme wrote: </div>
            </div>
            <blockquote type="cite">
              <div>
                <div class="h5">I didn't know about the theorem but it
                  seems reasonable to believe that some techniques are
                  more appropriate than others for modelling/predicting
                  financial data.<br>
                  It should be because of their nature : non gaussian,
                  non linear, non stationary, ...<br>
                  <br>
                  I think it's like the usual technical indicators
                  (MACD, Stochastic, etc ... ) which are more or less
                  suitable depending on the market conditions.<br>
                  <br>
                  What would be your recommended R package for NNs, as
                  there are different possible architecture : GRNN, PNN,
                  SOM ... (see <a moz-do-not-send="true"
                    href="http://en.wikipedia.org/wiki/NeuroSolutions"
                    target="_blank">http://en.wikipedia.org/wiki/NeuroSolutions</a>)<br>
                  <br>
                  Before trying many solutions, maybe it's worth to have
                  some discussion ... there can be another mighty
                  theorem we should all know :-)<br>
                  <br>
                  Best Rgds,<br>
                  <br>
                  Pierre<br>
                  <br>
                  <br>
                  <br>
                  <div class="gmail_quote">2011/8/23 Stephen Choularton
                    <span dir="ltr"><<a moz-do-not-send="true"
                        href="mailto:stephen@organicfoodmarkets.com.au"
                        target="_blank">stephen@organicfoodmarkets.com.au</a>></span><br>
                    <blockquote class="gmail_quote" style="margin:0pt
                      0pt 0pt 0.8ex;border-left:1px solid rgb(204, 204,
                      204);padding-left:1ex">I think that Gentil should
                      be aware of the /No Free Lunch Theorem/ (Duda et
                      al., 2001, Wolpert and Macready, 1997).  There are
                      no context-independent or usage-independent
                      reasons to favor one machine learning algorithm
                      over another. If one performs better than another,
                      it is owing to its better fit to the particular
                      problem, not its general superiority.  If you wish
                      to use these techniques try lots of them:
                       certainly neural networks and support vector
                      machines, but also try some of the ensemble
                      techniques such as bagging, boosting and random
                      forest.  You can even try the statisticians
                      favorite, logistic regression.  They are all
                      available in R.<br>
                      <font color="#888888"> <br>
                        Stephen Choularton Ph.D., FIoD</font>
                      <div>
                        <div><br>
                          <br>
                          <br>
                          On 23/08/2011 12:14 AM, Brian G. Peterson
                          wrote:<br>
                          <blockquote class="gmail_quote"
                            style="margin:0pt 0pt 0pt
                            0.8ex;border-left:1px solid rgb(204, 204,
                            204);padding-left:1ex"> On Mon, 2011-08-22
                            at 10:11 +0200, Gentil Homme wrote:<br>
                            <blockquote class="gmail_quote"
                              style="margin:0pt 0pt 0pt
                              0.8ex;border-left:1px solid rgb(204, 204,
                              204);padding-left:1ex"> I just send out
                              this post in order to share within
                              r-sig-finance any<br>
                              possible experience, advice, ... about NNs
                              or SVMs with R.<br>
                            </blockquote>
                            It seems that you're asking us to share with
                            you, and not sharing much<br>
                            yourself in return.<br>
                            <br>
                            Perhaps you could answer your own questions
                            in this thread with the<br>
                            things you are trying?<br>
                            <br>
                            SVM's have been discussed on this list many
                            times, please search the<br>
                            list archives.<br>
                            <br>
                            This blog has covered this topic:<br>
                            <a moz-do-not-send="true"
                              href="http://www.aphysicistinwallstreet.com/"
                              target="_blank">http://www.aphysicistinwallstreet.com/</a><br>
                            <br>
                            Also, there are a few books on machine
                            learning that use R.<br>
                            <br>
                            <blockquote class="gmail_quote"
                              style="margin:0pt 0pt 0pt
                              0.8ex;border-left:1px solid rgb(204, 204,
                              204);padding-left:1ex"> Several good
                              records have been published in the
                              litterature using these<br>
                              techniques for financial trading
                              strategies.<br>
                            </blockquote>
                            Which ones? References?<br>
                            <br>
                            <blockquote class="gmail_quote"
                              style="margin:0pt 0pt 0pt
                              0.8ex;border-left:1px solid rgb(204, 204,
                              204);padding-left:1ex"> There are also
                              commercial packages (expensive !) which
                              seem to have achieved<br>
                              good results.<br>
                            </blockquote>
                            Which packages?  References again?<br>
                            <br>
                            Note that neural network strategies are very
                            likely to create look-ahead<br>
                            bias as you develop and test them.  You try
                            something, fail, and try<br>
                            again on the same data.  Unless you are very
                            careful to reserve a 'pure'<br>
                            set of instruments and dates that you won't
                            *ever* touch until you think<br>
                            you have a 'good' machine learning system,
                            you're at pretty serious risk<br>
                            of introducing your look-ahead knowledge
                            into the system.  While this is<br>
                            true to one degree or another in any
                            quantitative strategy development,<br>
                            I think it is a particular risk in
                            self-adaptive machine learning<br>
                            methodologies.<br>
                            <br>
                            <br>
                            <blockquote class="gmail_quote"
                              style="margin:0pt 0pt 0pt
                              0.8ex;border-left:1px solid rgb(204, 204,
                              204);padding-left:1ex"> So I feel it could
                              be nice to share within this group about
                              the following<br>
                              subjects :<br>
                              <br>
                              - experience using the R packages<br>
                              - data pre-processing before feeding the
                              NNs (technical indicators,<br>
                              wavelets, EMDs, ....)<br>
                              - which type of NNs are suitable<br>
                              - how to build and train them<br>
                              - etc ...<br>
                              <br>
                              Thanks to all for sharing within the R
                              community<br>
                            </blockquote>
                            Now, your turn.  Bring the community up to
                            date with your research so<br>
                            far.<br>
                            <br>
                            Regards,<br>
                            <br>
                               - Brian<br>
                            <br>
                          </blockquote>
                          <br>
                        </div>
                      </div>
                      <div>
                        <div>
                          _______________________________________________<br>
                          <a moz-do-not-send="true"
                            href="mailto:R-SIG-Finance@r-project.org"
                            target="_blank">R-SIG-Finance@r-project.org</a>
                          mailing list<br>
                          <a moz-do-not-send="true"
                            href="https://stat.ethz.ch/mailman/listinfo/r-sig-finance"
                            target="_blank">https://stat.ethz.ch/mailman/listinfo/r-sig-finance</a><br>
                          -- Subscriber-posting only. If you want to
                          post, subscribe first.<br>
                          -- Also note that this is not the r-help list
                          where general R questions should go.<br>
                        </div>
                      </div>
                    </blockquote>
                  </div>
                  <br>
                </div>
              </div>
              <hr noshade="noshade" size="1">
              <p color="#000000" align="left">No virus found in this
                message.<br>
                Checked by AVG - <a moz-do-not-send="true"
                  href="http://www.avg.com" target="_blank">www.avg.com</a><br>
                Version: 10.0.1392 / Virus Database: 1520/3851 - Release
                Date: 08/22/11</p>
            </blockquote>
          </div>
          <br>
          _______________________________________________<br>
          <a moz-do-not-send="true"
            href="mailto:R-SIG-Finance@r-project.org">R-SIG-Finance@r-project.org</a>
          mailing list<br>
          <a moz-do-not-send="true"
            href="https://stat.ethz.ch/mailman/listinfo/r-sig-finance"
            target="_blank">https://stat.ethz.ch/mailman/listinfo/r-sig-finance</a><br>
          -- Subscriber-posting only. If you want to post, subscribe
          first.<br>
          -- Also note that this is not the r-help list where general R
          questions should go.<br>
        </blockquote>
      </div>
      <br>
      <hr noshade="noshade" size="1">
      <p class="avgcert" color="#000000" align="left">No virus found in
        this message.<br>
        Checked by AVG - <a moz-do-not-send="true"
          href="http://www.avg.com">www.avg.com</a><br>
        Version: 10.0.1392 / Virus Database: 1520/3851 - Release Date:
        08/22/11</p>
    </blockquote>
  </body>
</html>