Simple, but maybe difficult questions:1) What is a more interesting/succesfull approach to find patterns for trading: neural networks or hidden markov models? 2) What are the main diifferences between the two approaches? Are they just sophisticated moving averages?4) Which is the state of the art regarding these approaches?Thanks for your comments

QuoteOriginally posted by: JediWarriorSimple, but maybe difficult questions:1) What is a more interesting/succesfull approach to find patterns for trading: neural networks or hidden markov models? 2) What are the main diifferences between the two approaches? Are they just sophisticated moving averages?4) Which is the state of the art regarding these approaches?Thanks for your commentsPersonally I never worked with NN, so I can't say much about it. 1) Hard to say without significant evidence. I think HMM is very intuitive. The market usually work in states (bull/bear, high volalitity, etc) and markov switching is perfect for that.ck 2) ?? No, they arent moving averages. For HMM, check this:DUEKER, M., NEELY, C., J., Can Markov Switching Models Predict Excess ForeignExchange Returns? FRB of St. Louis Working Paper No. 2001-021F. Available at SSRN:http://ssrn.com/abstract=648227, 2001.HAMILTON, J., D. Regime Switching Models. Palgrave Dictionary of Economics, 2005.KIM, C., J., NELSON, C., R. State Space Model with Regime Switching: Classical andGibbs-Sampling Approaches with Applications. The MIT press, 1999.3) For NN I don't know, but for HMM, at least in econometrics, is markov switching state space models in a bayesian approach (gibbs sampler). But there are many simpler specifications.

After reading several papers about both techniques it seems that neural networks are more used for trading, meanwhile hidden markov is more for econometrics. Do you agree?

QuoteOriginally posted by: JediWarriorAfter reading several papers about both techniques it seems that neural networks are more used for trading, meanwhile hidden markov is more for econometrics. Do you agree?No I don't agree. My feeling is that there is a selection bias. You can find more papers on trading with NN because it looks good and at lot more people can do it (eg. download any package from the web).Btw, why you care?

QuoteOriginally posted by: msperlinQuoteOriginally posted by: JediWarriorMy feeling is that there is a selection bias. You can find more papers on trading with NN because it looks good and at lot more people can do it (eg. download any package from the web).There is no selection bias. NNs find application in trading because they are very good for forecasting if optimised properly, and not because NN package is easy to download. Some of the more solid works on NNs such as pi-sigma, ridge polynomial etc address issues HMMs fail to tackle.

QuoteOriginally posted by: muxControlQuoteOriginally posted by: msperlinQuoteOriginally posted by: JediWarriorMy feeling is that there is a selection bias. You can find more papers on trading with NN because it looks good and at lot more people can do it (eg. download any package from the web).There is no selection bias. NNs find application in trading because they are very good for forecasting if optimised properly, and not because NN package is easy to download. Some of the more solid works on NNs such as pi-sigma, ridge polynomial etc address issues HMMs fail to tackle.You could be right, but, again, we have no data to confirm it.

Hi,I think that one cannot say whether one method is better than another, it alldepend on the specific application. However, recently Support Vector Machine(SVM) was shown to generally give much better results than NNs. One can alsolook into combination of Machine Learning, Aritificial Intelligence methods likeNeuro-Fuzzy for e.g. One particular intersting aspect is to allow the user tobetter understand, e.g. get the trading rules, from the methods. Many paperspublished tend to use similar data set to allow fair comparisons on the basisof MSE, returns, number of transactions, etc. Brahim

I want to do a little research to find out an algorithm to help with my trading decisions. I am interested in intraday trading, of the market making type. Which method could be more relevant for this time frame?

QuoteOriginally posted by: msperlinMy feeling is that there is a selection bias. You can find more papers on trading with NN because it looks good and at lot more people can do it (eg. download any package from the web).Do you have data to confirm this?

QuoteOriginally posted by: muxControlQuoteOriginally posted by: msperlinMy feeling is that there is a selection bias. You can find more papers on trading with NN because it looks good and at lot more people can do it (eg. download any package from the web).Do you have data to confirm this?QuoteYou could be right, but, again, we have no data to confirm it.As I've said before, no, we don't have data to confirm either way.

QuoteOriginally posted by: JediWarriorI want to do a little research to find out an algorithm to help with my trading decisions. I am interested in intraday trading, of the market making type. Which method could be more relevant for this time frame?You're going to have to figure out this one on your own...Don't expect people to open up their algo to you..

QuoteMany papers published tend to use similar data set to allow fair comparisons on the basisof MSE, returns, number of transactions, etc. And you see this as a good thing ?Unless their data is very long and very big (eg. 100+ assets) there is a great potential for data snooping...

QuoteOriginally posted by: msperlinQuoteMany papers published tend to use similar data set to allow fair comparisons on the basisof MSE, returns, number of transactions, etc. And you see this as a good thing ?Unless their data is very long and very big (eg. 100+ assets) there is a great potential for data snooping...Are you saying researchers are liars?.... If so then it confirms my doubts about some of your papers.... now I know why you submitted to ssrn and not Elsevier/IEEE.

QuoteOriginally posted by: muxControlQuoteOriginally posted by: msperlinQuoteMany papers published tend to use similar data set to allow fair comparisons on the basisof MSE, returns, number of transactions, etc. And you see this as a good thing ?Unless their data is very long and very big (eg. 100+ assets) there is a great potential for data snooping...Are you saying researchers are liars?.... If so then it confirms my doubts about some of your papers.... now I know why you submitted to ssrn and not Elsevier/IEEE.liars ? why ?data snoopingLighten up man. There is no reason for getting personal.If you have doubts about my papers (any of them), lets hear it.

Hi,What I mean about performance evaluation was that many scientific/academic papersuse specific measure which may not be seen as appropriate by experts. MSE is a poormeasure I think. For e.g. in audio / image / video quality measures such as Signal toNoise (SNR) ratio was used until now but perceptual index have been developedwhich are more found objective and robust. Maybe some similar measures could new research topics for research in finance. Surely statistical significance (as beingdiscussed in one other forum) could be one. Maybe some experiences members couldgive some input on this. I do remember a post by someone detailing what he thoughtwould be a good methodology to assess a trading strategy.I agree with msperlin, there indeed no need to become personal. His code isfreely available and allowing people to replicate the results he presented inhis papers. As for publishing standards with SSRN vs Elsevier vs IEEE, I amsure his work would be publishable in some Elsevier journals. I do not knowthe SSRN review process but papers are of very good quality I think (my ownview).Let's have some constructive discussion and avoid being trapped into silly arguments. talking about data, are there any competition/contest on financialdata set like the ones organised in some conference on bioinformatics orpattern recognition ?Brahim