Multi time sereis

so
NiCd Battery

Multi time sereis

Hi,
I am training multi time series model and  data contains spikes. Does anyone know how to handle the spikes in DataRobot?

Thankyou!

Labels (1)
0 Kudos
6 Replies

not an expert! but the Improving Time Series Models webinar I watched last week might help you https://community.datarobot.com/t5/sessions/improving-time-series-models/ba-p/5742

 

this is in the attached pdf - 

"Certain clusters have distinct time trends, for example only Cluster 1 (orange) has a sharp sales spike
in August each year. This is strong indication you may want to build separate projects per cluster" - fyi

0 Kudos

Thanks for your kind response!  Actually my data has more granularity. for example  one cluster contains further time series containing random spikes. Its not fixed.

0 Kudos
Anatolii
DataRobot Employee
DataRobot Employee

Hi @so ,

 

In general, if your spikes can be explained by the variables you have provided in the input data, DataRobot models should provide predictions that will cover them accurately.

 

If these spikes are caused by special events, like sales or holidays, you can add a calendar file to your Time Series project. With it, DataRobot will generate special features which will help model predicting such special event spikes accurately.

 

As per @sam632 recommendation, spikes might be a characteristic of a certain cluster of your series. In this case, try modeling this cluster in a separate project. This way a model will have more chances to find a useful pattern in historical data to forecast such spikes.

0 Kudos

Thanks you!

0 Kudos

Hey @so Looks like @Anatolii answered your question for you? If that was the help you were looking for, can you please accept his reply as the solution? This will help other community members get help too! thanks for the housekeeping task 🙂

Linda

0 Kudos

Hi,
@Linda @Anatolii thanks a lot for your suggestions. It was really helpful in exploring the data. Unfortunately, I could not get the solution due to changing nature of the data..
Thank you!

0 Kudos