[Neural Forecasting] approach to design and training for times series forecasting using ANN

Hi All



I am currently doing a college project on forecasting using
ANNs. At the moment I am trying to build a sample time series using JOONE.


 


For the following time series example 0.1, 0.2, 0.3, 0.4, 0.5,
0.6, 0.7, 0.8, 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1 (just a simple up
and down spike) I was wondering what the best approach to a design and training
would be.


 


At the moment my thoughts are using - 1 inputs in input layer, 15 neurons in 1 
hidden layers, 1 output in output layer for the design and for training take the
first input and set the next value as the desired output. For example 0.1 as
input set 0.2 as the desired output, 0.2 set 0.3 as desired output etc…


 


My second approach is using - 17 inputs in input layer, 51/68
neurons in 1 hidden layers, 1 output in output layer for the design and for 
training take the whole
series and take the next value as the desired output, like a sliding window.


For example for the series 0.1, 0.2, 0.3, 0.4, 0.5, 0.6,
0.7, 0.8, 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1 set the desired output to
0.1


for the series 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.8,
0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1,0.1 set the desired output to 0.3


for the series 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.8, 0.7,
0.6, 0.5, 0.4, 0.3, 0.2, 0.1,0.1,0.2 set the desired output to 0.3 etc…

I am using sigmoid for the transfer functions.



 


Any suggestions, flaws or better approaches would be greatly
appreciated.


 


Cheers,
Eamon
 
Eamon Doyle
eamon121@xxxxxxxx
Tel: 087-2374491 
check out my online lotto syndicate http://www.v-w-d.com/156964ED
the smarter way to play




Send instant messages to your online friends http://uk.messenger.yahoo.com

Other related posts: