Help Required:- Forecast in Machine learning is not working as expected

Hi Team,

I Tried using machine learning for anomaly detection & prediction using forecast feature.... I have loaded 1 year events(alarms) data & trying to do prediction of alert count by splitting domains(domain wise alert prediction using multi metric option to achieve this)

When I tried to forecast this data, predicted graph is showing linear(straight line)

I have tried similar option for CPU Utilization prediction & it is working perfectly... How can I rectify this. Why the forecasted data shows straight line for next N Days, when there is a so many fluctuation in historical data.

Hi Shridhar,

from the screenshot you provided I see on the preview at the bottom that there was a step change happening around mid january. Before that change the data seemed to be rather flat somewhat on the level of the forecast. It therefore seems to me that the model did not adjust after the step change and it also did not find any repeated pattern, the spikes look somewhat arbitrary.

What you can do:

  1. You can try cloning the job and start feeding data after the step change mid january omitting everything before.

  2. You can enable model plots to get a visualization of the modeling process by adding the following to the job config:

    "model_plot_config" : {
    "enabled" : true
    }

I hope this helps. We are working on further improvements of forecasting.

Hi Hendrik,

Thanks for your inputs.. I will check this and get back to you...

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