Prediction of euro exchange rate

[Misa]

Hi everyone,

have someone experience with prediction next exchange rate number from ten previous values?

I have training set where for every ten exchange rates (as data) is label with one following value. In data are ten values shifting by one place - first value is take away and in the end is add following value from exchange rate raw.

Data Example
 
data:
                             10         9         8         7         6         5          4         3         2          1

A  28.40999985  | 28.29000092 |  [ 28.05999947  28.05999947  28.05999947  27.86499977  27.78499985  27.69499969  27.67499924  27.70000076  27.70000076  27.70000076]

B  28.22999954  | 28.40999985 |  [ 28.29000092  28.05999947  28.05999947  28.05999947  27.86499977  27.78499985  27.69499969  27.67499924  27.70000076  27.70000076]  27.70000076

C  27.84000015  | 28.22999954 |  [ 28.40999985  28.29000092  28.05999947  28.05999947  28.05999947  27.86499977  27.78499985  27.69499969  27.67499924  27.70000076]  27.70000076  27.70000076

                  v
labels:
         A                   B                  C
[ 28.29000092  28.40999985  28.22999954 ]
 

First question is, if it is okay that the data are generating from right to left or it should be problem in training phase?

And other question is if my net design is right for training this network:

Through this network I got the best Euclidean loss 0.590581.

When I add this caffemodel to predict some exchange ratings I don't know which layer I should add to get predicted number.My predicting model looks like this:

When I print result from 'fc3' Blob I get: {'fc3': array([[ 28.1656723]], dtype=float32)} for all input values. I would like to ask what the printed result means and how to change net configuration to get better results.

Thanks for help!

[Jan]

Completely disregarding your actual question I would like to suggest you to use a recurrent network approach, as this sounds like a classic sequence learning task. I don't think caffe can handle recurrency at the moment, but there is a PR that gives hope that LSTM and RNN layers may soon be possible (you can even merge it yourself and try it today): https://github.com/BVLC/caffe/pull/3948

Jan