Forecasting using SAS Forecast Server, SAS/ETS, and more

Modelling ARIMA-ANN

Accepted Solution Solved
Reply
Highlighted
Occasional Contributor
Posts: 5
Accepted Solution

Modelling ARIMA-ANN

Hello Pals;

 

Can anyone give me an idea of how to model auto regressive artificial neural network (ARIMA-ANN) in SAS UE, using the data below.

 

data WORK.WINE;
    infile datalines dsd truncover;
    input date:$8. y:32.;
    datalines4;
Jan80,15136
Feb80,16733
Mar80,20016
Apr80,17708
May80,18019
Jun80,19227
Jul80,22893
Aug80,23739
Sep80,21133
Oct80,22591
Nov80,26786
Dec80,29740
Jan81,15028
Feb81,17977
Mar81,20008
Apr81,21354
May81,19498
Jun81,22125
Jul81,25817
Aug81,28779
Sep81,20960
Oct81,22254
Nov81,27392
Dec81,29945
Jan82,16933
Feb82,17892
Mar82,20533
Apr82,23569
May82,22417
Jun82,22084
Jul82,26580
Aug82,27454
Sep82,24081
Oct82,23451
Nov82,28991
Dec82,31386
Jan83,16896
Feb83,20045
Mar83,23471
Apr83,21747
May83,25621
Jun83,23859
Jul83,25500
Aug83,30998
Sep83,24475
Oct83,23145
Nov83,29701
Dec83,34365
Jan84,17556
Feb84,22077
Mar84,25702
Apr84,22214
May84,26886
Jun84,23191
Jul84,27831
Aug84,35406
Sep84,23195
Oct84,25110
Nov84,30009
Dec84,36242
Jan85,18450
Feb85,21845
Mar85,26488
Apr85,22394
May85,28057
Jun85,25451
Jul85,24872
Aug85,33424
Sep85,24052
Oct85,28449
Nov85,33533
Dec85,37351
Jan86,19969
Feb86,21701
Mar86,26249
Apr86,24493
May86,24603
Jun86,26485
Jul86,30723
Aug86,34569
Sep86,26689
Oct86,26157
Nov86,32064
Dec86,38870
Jan87,21337
Feb87,19419
Mar87,23166
Apr87,28286
May87,24570
Jun87,24001
Jul87,33151
Aug87,24878
Sep87,26804
Oct87,28967
Nov87,33311
Dec87,40226
Jan88,20504
Feb88,23060
Mar88,23562
Apr88,27562
May88,23940
Jun88,24584
Jul88,34303
Aug88,25517
Sep88,23494
Oct88,29095
Nov88,32903
Dec88,34379
Jan89,16991
Feb89,21109
Mar89,23740
Apr89,25552
May89,21752
Jun89,20294
Jul89,29009
Aug89,25500
Sep89,24166
Oct89,26960
Nov89,31222
Dec89,38641
Jan90,14672
Feb90,17543
Mar90,25453
Apr90,32683
May90,22449
Jun90,22316
Jul90,27595
Aug90,25451
Sep90,25421
Oct90,25288
Nov90,32568
Dec90,35110
Jan91,16052
Feb91,22146
Mar91,21198
Apr91,19543
May91,22084
Jun91,23816
Jul91,29961
Aug91,26773
Sep91,26635
Oct91,26972
Nov91,30207
Dec91,38687
Jan92,16974
Feb92,21697
Mar92,24179
Apr92,23757
May92,25013
Jun92,24019
Jul92,30345
Aug92,24488
Sep92,25156
Oct92,25650
Nov92,30923
Dec92,37240
Jan93,17466
Feb93,19463
Mar93,24352
Apr93,26805
May93,25236
Jun93,24735
Jul93,29356
Aug93,31234
Sep93,22724
Oct93,28496
Nov93,32857
Dec93,37198
Jan94,13652
Feb94,22784
Mar94,23565
Apr94,26323
May94,23779
Jun94,27549
Jul94,29660
Aug94,23356
;;;;


Accepted Solutions
Solution
‎11-17-2017 09:14 AM
SAS Super FREQ
Posts: 100

Re: Hybrid model issue

There are several ways to model an "ANN" model.

 

1. Using neural net alone to model ANN(p,k) (i think it should be p, not q). You have to create p lags of the dependent variable as the input to the neural net with k hidden neurons. In addition, you can diff the data first and then modeling the diff instead of the original series so you are modeling something like ANN(p,d,k)

 

2. Stacking models. Take the residuals from the ANN model from 1 and modeling them using ARIMA(p,d,q). The final forecast will be the ANN forecast + ARIMA forecast

 

3. Ensemble models. Model ANN and ARIMA separately and take some sort of averaging (e.g. straight up average or regression based averaging) of the forecasts

 

4. Recurrent neural net. This model is much more complex and takes a lot of time to train. In practice methods 1 to 3 should be more than sufficient.

thanks

Alex

View solution in original post


All Replies
Super User
Posts: 20,696

Re: Modelling ARIMA-ANN

Unfortunately that type of model is not supported within SAS UE Smiley Sad

 

SAS UE does offer ARIMA but not neural networks, as far as I know. 

 

 

Occasional Contributor
Posts: 5

Re: Modelling ARIMA-ANN

Thanks,
but what suggestion can you give me. I am using it for my Thesis
Super User
Posts: 20,696

Re: Modelling ARIMA-ANN

Check if your University offers the full version of SAS - its usually available for a small cost, $99 at my University. 

 

Or choose another software package such as Python or R. 

Occasional Contributor
Posts: 5

Hybrid model issue

[ Edited ]

Hello Pals;

 

 Can someone give me an idea of how to model an ARIMA(p,d,q)+ANN(q,k) 'autoregressive Neural network model in SAS UE, with the data below:-

 

data WORK.WINE; 
  infile datalines dsd truncover; 
  input date:$8. y:32.; 
datalines4; 
Jan80,15136 
Feb80,16733 
Mar80,20016 
Apr80,17708 
May80,18019 
Jun80,19227 
Jul80,22893 
Aug80,23739 
Sep80,21133 
Oct80,22591 
Nov80,26786 
SAS Employee
Posts: 9

Re: Hybrid model issue

As mentioned above, neural networks are not supported in SAS University Edition. However, the general approach would be to model the data series with ARIMA for linear effects and then fit a neural network model to the residuals for non-linear effects. 

 

You can use University Edition for the ARIMA model, but you will either have to get access to a different version of SAS that includes neural networks or use a different programming environment for the second part.

Occasional Contributor
Posts: 5

Re: Hybrid model issue

Posted in reply to solarflare

Thank you

Solution
‎11-17-2017 09:14 AM
SAS Super FREQ
Posts: 100

Re: Hybrid model issue

There are several ways to model an "ANN" model.

 

1. Using neural net alone to model ANN(p,k) (i think it should be p, not q). You have to create p lags of the dependent variable as the input to the neural net with k hidden neurons. In addition, you can diff the data first and then modeling the diff instead of the original series so you are modeling something like ANN(p,d,k)

 

2. Stacking models. Take the residuals from the ANN model from 1 and modeling them using ARIMA(p,d,q). The final forecast will be the ANN forecast + ARIMA forecast

 

3. Ensemble models. Model ANN and ARIMA separately and take some sort of averaging (e.g. straight up average or regression based averaging) of the forecasts

 

4. Recurrent neural net. This model is much more complex and takes a lot of time to train. In practice methods 1 to 3 should be more than sufficient.

thanks

Alex

SAS Employee
Posts: 7

Re: Modelling ARIMA-ANN

You can solve it by using Time Delay Neural Networks using PROC NEURAL. You have to create a skip layer perceptron architecture, having input nodes connected to hidden nodes and connected direct to the output node. 

☑ This topic is solved.

Need further help from the community? Please ask a new question.

Discussion stats
  • 8 replies
  • 488 views
  • 1 like
  • 5 in conversation