ML monthly forecast

machine-learning

(Felix Werkmeister) #1

Hi community,

I just started to do some testing with X-Pack machine learning and especially the forecasting feature. I have very simple data with (time, value) on a monthly basis with 60 datapoints (5 years) in history. Until now I did not achieve a very good result. I choose different buckets (between 30d and 365d) as well as different aggregations.

  • Does somebody have a similar usecase and made some experience?
  • Is this a usecase that would be feasible?

Thanks a lot,
best regards,

Felix


(rich collier) #2

Hello Felix,

In general, you probably simply don't have enough data. Only having 60 points is not enough to create a robust statistical model of data - having on the order of hundreds of data points should be sufficient, however. Do you have access to daily (or hourly) data instead?


(Felix Werkmeister) #3

Hello Rich,

thank's for the fast feedback. Sadly I only have the monthly data and it's true that 60 points are not very much. We allready experienced a bit with ARIMA and LLP models (programing it manually) and it worked not pwerfect but quite ok although there are only 60 points. Do you know what kind of model is used within the ML Library?
I also recognized that the forecast is only possible for the next 8 weeks. Is it possible to change it to the next x points in future?

Thanks and regards,
Felix


(Hendrik Muhs) #4

Hello Felix,

Forecast is using the very same models that are used for anomaly detection. Hardly to explain here, it is a bunch of different algorithms, you can get an idea watching last years ElasticON talk:
https://www.elastic.co/elasticon/conf/2017/sf/machine-learning-and-statistical-methods-for-time-series-analysis

Regarding the forecast window: It is unfortunately not possible to forecast beyond the limit of 8 weeks, it is fixed for the current version.

Having that said we are working on improving forecasting so that we can change the limit in an upcoming version. Your feedback is appreciated.


(Felix Werkmeister) #5

Hi Hendrik,

Thanks for the info and the video link. Then I will leave it for now and try again with the upcoming versions and also try to get more data.

Regards,
Felix


(system) #6

This topic was automatically closed 28 days after the last reply. New replies are no longer allowed.