1
$\begingroup$

I am playing with time series data related to a issue ticketing system. The system logs all open tickets at any one point and my task is to predict what the volume of open tickets will be in 5,10,15 days in the future. I was thinking of avoiding the stereotypical ARIMA style models for a machine learning approach with a lot of feature engineering as the data has a lot of powerful associated information.

My plan was to build a testing data set of my predictor variables (#open tickets, 10 day avg close rate, 10 day avg open rate, profile of the open tickets, profile of there close rate etc.) and use them to predict #open tickets 5 days into the future with traditional algorithms(NN,RF etc).

Just wondering is the approach a good one and general pros and cons.

Improve this question
$\endgroup$
2
  • $\begingroup$ There is an extension of ARIMA models to Transfer Function models which include detection of optimal lead and lag effects of both use-specified series and holiday effects. This is all done with an eye for necessary augmentations like Pulses,Level Shifts,Seasonal Pulses, Local Time Trends and possible time varying parameters/ time varying error variance components. If you want to post your data in an excel file I will be glad to provide a potential road-map for you. $\endgroup$
    – IrishStat
    Commented May 4, 2016 at 16:39
  • $\begingroup$ Something to keep in mind: you're dealing with counts (counts tend to have some special properties, like being non-negative but with exact zeros, heteroskedastic, necessarily nonlinear relationships because of the 0 boundary, ...); there are models suitable for counts $\endgroup$
    – Glen_b
    Commented May 5, 2016 at 0:58

1 Answer 1

Reset to default
1
$\begingroup$

This is certainly a valid approach. Just be sure to include dummies to model "seasonal" behavior - if your product ships every six months and you do lots of integration tests every six months, I'd expect a spike in tickets around that time, so be sure your ML algorithms have an understanding of time and can pick up on this half-yearly seasonality. (Or even better, if this is driven by external factors like integration testing, feed this to your algorithms.)

Sven Crone, currently at Lancaster Business School, used to do forecasting with Neural Networks, though I haven't seen a lot of recent work from him on this - might be a good idea to skim some papers of his. Here are pointers to using Random Forests for forecasting. Finally, you may want to search through earlier questions tagged both "forecasting" and "random-forest" (modify using "prediction" instead of "forecasting", or other ML algorithm tags), although this was surprisingly unhelpful.

Improve this answer
$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.