Explorer==>Forecast==>Specifying Lag Length

classic Classic list List threaded Threaded
2 messages Options
Reply | Threaded
Open this post in threaded view
|

Explorer==>Forecast==>Specifying Lag Length

Chaitanya Birudavolu
Hi,

I am referring to WekaExplorer==>ForecastPanel==>AdvancedConfigrationPanel==>LagCreation==>Panel.

There are 3 fields that control the lag creation, but I'm wondering why that is designed so.

The three fields are "Maximum lag", "Minimum lag", and "Fine tune lag selection".

Perhaps it could have been designed as just a single field where the user can enter a comma-separated list of lag-lengths -- for example, 3,6-8 (means we want lag variables (current-3), (current-6), (current--7), and (current-8).

I have a related observation to report:

Suppose I load airline.arff, (which has 144 instances) and enter the following settings:
BasicConfiguration==>BasicConfiguration==>Periodicity=Monthly
Perform Evaluation=On
AdvancedConfiguration==>MinimumLagLength=1
AdvancedConfiguration==>MaximumLagLength=12
AdvancedConfiguration==>FinTuenLagSelection=3

With these settings, it should be clear to Weka that we do not want lag-length to exceed 3.
Hence, while performing evaluation, we should expect metrics for one-step-ahead predictions to be evaluated at 141 points in the series (144 minus 3 = 141).

But I find that the one-step-ahead predictions are evaluated only at 132 points (144 minus 12 = 132).

Clearly, when it comes to the question of how many evaluations to perform, Weka is looking at the setting MaximumLagLength=12. but it is ignoring the setting "FineTuneLagSelection"=3

I see no reason why the evaluation should happen only 132 points in the data series. Given that we have fine-tuned to have a maximum lag-length of 3, we will have lag-variables only for (current-1), (current-2), (current-3), which means that we will have missing values only in the first 3 rows of the data-series.

Could someone please let me know if I'm missing something in my understanding -- either or Weka, or of how the Time Series Analysis works.

Cheers,
Chaitanya
_______________________________________________
Wekalist mailing list
Send posts to: [hidden email]
To subscribe, unsubscribe, etc., visit https://list.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html
Reply | Threaded
Open this post in threaded view
|

Re: Explorer==>Forecast==>Specifying Lag Length

Mark Hall
Yes, I guess this is a little unintuitive, and it should take the fine tune settings into account when determining which points to produce predictions for. Having the minimum and maximum (most of the time folks just use maximum) settings was an attempt at making (an already feature rich system) slightly less complicated. The fine-tune settings are for those who like to fiddle at bit more :-) Note that you can get the result you are looking for by setting the maximum lag length to 3.

Cheers,
Mark.

´╗┐On 18/11/18, 12:42 AM, "Chaitanya Birudavolu" <[hidden email] on behalf of [hidden email]> wrote:

    Hi,
   
    I am referring to WekaExplorer==>ForecastPanel==>AdvancedConfigrationPanel==>LagCreation==>Panel.
   
    There are 3 fields that control the lag creation, but I'm wondering why that is designed so.
   
    The three fields are "Maximum lag", "Minimum lag", and "Fine tune lag selection".
   
    Perhaps it could have been designed as just a single field where the user can enter a comma-separated list of lag-lengths -- for example, 3,6-8 (means we want lag variables (current-3), (current-6), (current--7), and (current-8).
   
    I have a related observation to report:
   
    Suppose I load airline.arff, (which has 144 instances) and enter the following settings:
    BasicConfiguration==>BasicConfiguration==>Periodicity=Monthly
    Perform Evaluation=On
    AdvancedConfiguration==>MinimumLagLength=1
    AdvancedConfiguration==>MaximumLagLength=12
    AdvancedConfiguration==>FinTuenLagSelection=3
   
    With these settings, it should be clear to Weka that we do not want lag-length to exceed 3.
    Hence, while performing evaluation, we should expect metrics for one-step-ahead predictions to be evaluated at 141 points in the series (144 minus 3 = 141).
   
    But I find that the one-step-ahead predictions are evaluated only at 132 points (144 minus 12 = 132).
   
    Clearly, when it comes to the question of how many evaluations to perform, Weka is looking at the setting MaximumLagLength=12. but it is ignoring the setting "FineTuneLagSelection"=3
   
    I see no reason why the evaluation should happen only 132 points in the data series. Given that we have fine-tuned to have a maximum lag-length of 3, we will have lag-variables only for (current-1), (current-2), (current-3), which means that we will have missing values only in the first 3 rows of the data-series.
   
    Could someone please let me know if I'm missing something in my understanding -- either or Weka, or of how the Time Series Analysis works.
   
    Cheers,
    Chaitanya
    _______________________________________________
    Wekalist mailing list
    Send posts to: [hidden email]
    To subscribe, unsubscribe, etc., visit https://list.waikato.ac.nz/mailman/listinfo/wekalist
    List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html
   


_______________________________________________
Wekalist mailing list
Send posts to: [hidden email]
To subscribe, unsubscribe, etc., visit https://list.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html