straight line using AutoArima for time series forecasting!

Options
AVerhagen
AVerhagen Dataiku DSS Core Designer, Dataiku DSS ML Practitioner, Registered Posts: 6

Hi,

I'm currently working on a time series forecast using auto ARIMA. My dataset contains weekly peaks and I need to forecast one year ahead. However, when I use auto ARIMA, the resulting forecast appears as a straigth line without capturing seasonality and patterns that are evident in the historical data. The historical data spans from 2017 to 2023, featuring weekly peaks and an increasing trend each year.

I have attempted to modify the preset parameters, such as adjusting the season length to values like 4 weeks or 52 weeks and D=1. Unfortunately, these changes still yield a forecast that lacks the desired pattern and fails to show the identified seasonality and trend.

What specific parameters should I modify? How can I achieve a forecast that accurately reflects the seasonality and trend present in my dataset? Any insights or advice would help! Thanks in advance

AutoARIMA settings 1.png

AutoARIMA settings 2.png

Thanks in advance!

Answers

  • StanG
    StanG Dataiker, Registered Posts: 52 Dataiker
    Options

    Hi,

    If you get a straight line, it's probably because the search for the best arima orders failed to converge due to a too large search space and a limited number of search iterations. Sometimes removing the seasonal component of arima (setting the season length to 1) can help the search converge by reducing the search space and might improve a lot the forecasts.

    Also, are you forecasting weekly values ? And is your horizon 52 ?

  • AVerhagen
    AVerhagen Dataiku DSS Core Designer, Dataiku DSS ML Practitioner, Registered Posts: 6
    Options

    Hi, thanks for your reply!

    I also tried season length = 1 but it still shows a straigth line.
    Yes I'm forecasting weekly values and my forecast horizon is 52 weeks.

    Could it be that I have to set a frequency parameter and how do I implement this? It doesn't seem to pick up any patterns or trends when I vary the season length.

  • StanG
    StanG Dataiker, Registered Posts: 52 Dataiker
    Options

    What version of DSS are you using ? And what is the version of pmdarima in your code env ?

  • AVerhagen
    AVerhagen Dataiku DSS Core Designer, Dataiku DSS ML Practitioner, Registered Posts: 6
    Options

    I'm using DSS 11 and pmdarima==1.2.1

  • StanG
    StanG Dataiker, Registered Posts: 52 Dataiker
    Options

    Ok thanks

    I can recommend that you update DSS to version 12 that supports pmdarima up to 1.8.4. Updating pmdarima might improve the results and also starting from DSS 12 you can play with the parameter maxiter to increase the arima search.

    If you cannot update DSS, you can try to just update pmdarima even though it won't be supported (I think that the maximum version that will work on DSS 11 is 1.7).

  • AVerhagen
    AVerhagen Dataiku DSS Core Designer, Dataiku DSS ML Practitioner, Registered Posts: 6
    Options

    This doesn't seem to make any difference. Thanks for trying to help!

Setup Info
    Tags
      Help me…