Does advanced functional analysis make any sense?

I decided to do a home project.
What I did can be expressed in 3 simple steps:

1. I disassembled markat data into separate waves.

Screenshot1.png

2. Predict 5 new possible readings using arima.

Screenshot2.png

3. Reconstract market data from the new readings.

So here's what I've got:
Before you watch, bear in mind that:
There are 5 lines because I was switching between buffers because I didn't want to overwrite previous prediction with the new one.
The predictions are done for Open prices.
When you see a line staring at bar X, it means that the prediction was calculated on previous bar opening.
The prediction in this case was based on sample of 60 bars.
Screenshot3.png

So I have presented to you the results of my little experiment.

As far as I'm concerned This doesn't seem to work, at least not for small samples. Maybe if I had a cluster of 1440 CPU's and the sample would consist of data from a whole week, the results would be better, but who knows.

Please, if anybody has any experiance or any other observations he/she would like to share, please do.
I look forward to your comments.

I tried wavelet decomposition. I also had a look at the empirical mode decomposition (there is a R module).

60 bars are way too small a sample but you should be able to cruch a week worth of data without involving the amazon cluster.
The real problem is always the same: the rightmost end of the chart. Can you draw the interval of confidence as well? You'll see that you're certainly inside. And that they are very big :-(
Sorry I wish I had a better feedback.

Nothing to be sorry about. In your short post you gave me at least two clues and directions worth checking.

You said that week of data shouldn't be required.
Well I use auto.arima from package 'forecast' and I have tell you that it takes almost 60 seconds to forcast 60 frequencies using this function. In your research, did you stumble upon any other fitting/extrapolating algorithm that would work comparatively well?

Ensembles always work better than single estimators. There is a lot of research about different estimators but the problem always almost (especially eith trading IMHO) comes to how you build the feature set for your regressor/classifier.

Here is a good paper with comparison of multiple algorithms (unfortunately extremely randomized forests are not included) - page 10 ->

http://citeseerx.ist.psu.edu/viewdoc...=rep1&type=pdf

this paper is also worth reading

http://orbi.ulg.ac.be/bitstream/2268...lj-advance.pdf