How is Time Series Forecasting with Neural Networks Implemented in Weka?

[Erick Butler Poletto]

I am taking both this course and the Machine Learning Coursera one to have the necessary knowledge to implement an electronics price forecasting model.

I have several months of retail prices for several products and from the research I have done (see N.K. Ahmed, An empirical comparison of machine learning models for time series forecasting) , the multilayer perceptron is one of the best predicting methods for time series.

The Weka implementation seems very robust for prototyping but neither in its documentation or in general literature I could find insights on how to implement:

1) A neural network that forecasts one individual time series. From what I understood it would take the last m observations (in weeks?days?) as its input. The hidden layer would apply logistic regression (the number of layers/nodes is not clear) and the final layer/node would apply linear regression.

2) A neural network that forecasts one time series based on several time series (as can be done in Weka). The reason to do this would be to consider foreign exchange time series and other brand prices on the price of a given product. Is this extra input considered in the 1st layer or in the hidden layers? How relevant time series can be identifyed (even w/ time lag)?

3) How to model punctual events (holidays, new product release dates) in the neural network? Should these factors be considered outside the model?

Any insight on the weka implementation would be of great help! I look forward using it to test the soundness of the forecasts for my datasets but I would really like to understand how the calculations are done to implement them myself.

Thanks!

[Fernando DM]

Hi Erick.

For more information, I understand that neural networks is one of the last topics of the course (Data Mining with WEKA), however:

For the Multilayer perceptron architecture, it is necessary to determine the structure and the weights of the network, there are several ways: 

Testing, testing, and testing, although there are approaches using algorithms, e.g. Backpropagation 

Some tips for networks are: (are heuristics) 

- Changing the order of the patterns randomly. 

- Check the types of data (if you have different scales), because it could affect the weights 

- Check values of alpha. 

- Test with random weights 

About the Layers: In the manual of "WEKA" you can find: 

Hiddenlayers -- This defines the hidden layers of the neural network. This is a list of positive whole numbers. 

1 For each hidden layer. Comma Seperated. To have no hidden layers put a single 0 here. This will only be used if Autobuild is set. There are also wildcard values 'a' = (attribs classes) / 2, 'i' = attribs, 'o' = classes , 't' = attribs classes.

The number of hidden layers depends on the problem and can be determined, if the network is larger than necessary for the problem (the network could find local minimum)

In "WEKA" you can create variables, based on other variables, i think you refer to that when you talk about variables of anniversary (holidays, new product release dates), and that item, is also addressed in the following lessons. Of course, you should be clear which is the time event that you want to model.

Best Regards

Ing. Fernando García

Data Mining Specialist

ar.linkedin.com/in/fernand0garcia/

[Fernando DM]

Hi Erick, 
I just checked and I do not think that the mooc covers (as you need) many of the points you mention. 

However, you can check the following comment from IAn:

"https://groups.google.com/forum/#!topic/wekamooc-general/7ua9IjvGdHc"

Best Regards

Ing. Fernando García

Data Mining Specialist

ar.linkedin.com/in/fernand0garcia/

El martes, 4 de marzo de 2014 18:14:42 UTC-3, Erick Butler Poletto escribió:

[Erick Butler Poletto]

Hi Fernando,

Thanks for the pointers, I will check the other post for more information!

Best,

Erick