1
$\begingroup$

I have a conceptual question and not sure what would be the appropriate solution.

I have run a time series forecast using arima methodology. I had several years of data that I used and split my data into training and validation, and proceeded to train the model using the training portion and did a forecast on the validation data in order to evaluate the efficiency of my forecast. I have also done the preprocessing steps to clean the dataset for the outliers before I ran the model.

However, when I look at the results of the forecast in the validation set, I see that the actuals significantly exceeded the forecast for a particular date. Upon further research, I uncovered that there was a marketing campaign that was run at that time that was never before run in all the historical data for that date.

I tried different approaches using dummy variables, and somehow I am not able to figure out how I can bring my forecast to do a better job for this unseen value in the validation data that is never before seen in the training data.

Any feedback or alternative approaches to this scenario would be much appreciated. Thanks!

$\endgroup$
7
  • $\begingroup$ Is extreme value also seen in test data? $\endgroup$
    – forecaster
    Dec 13, 2018 at 0:32
  • 1
    $\begingroup$ @forecaster, Thanks for your response. The extreme value is seen only in the test data, but not in the training data.The response variable (in this case, sales) saw a significant spike for that week in the validation(test) data, but in the historical data, that type of spike was not seen for the same comparable week.I also want to add that I am using the forecast package in R from Dr. Hyndman. $\endgroup$
    – iRm
    Dec 13, 2018 at 14:42
  • $\begingroup$ methods that use historical data to model can earn only when there is event present in test data, you cannot expect the model to identify unknown or unexpected events that it has not “seen” in the past. $\endgroup$
    – forecaster
    Dec 13, 2018 at 15:02
  • $\begingroup$ The event in question is Easter and a campaign was run the prior week. It does fall on different weeks each year and I used dummy variables to account for this. But, the sales jump in the validation data is much higher than in the historical periods. So, just to confirm, even if the event is known, the significant spike in the validation data that is not seen in the historical data can explain why the forecast did poorly this week? Thanks! $\endgroup$
    – iRm
    Dec 13, 2018 at 15:15
  • $\begingroup$ Are there any other methods that can be used for this? $\endgroup$
    – iRm
    Dec 13, 2018 at 15:16

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.