Forecasting Multiple Time Series with Modeltime | Bonus Auto-Forecast Shiny App [Lab 46]

[@filipposstamatelos9565]

In my opinion, this is a great lab. Also, your shinny app is amazing! That gives many ideas and the modeltime itself saves really a lot of time

[@BusinessScience]

Excellent. I'm glad to hear it. Forecasting at scale gives rise to many possibilities. Auto-forecasting is just one of many tools that you can build.

[@sebastiansantiago3783]

Hi Matt…your content is great…if model time use forecast package…need to check fabeel is the next version of forecast 🙏🏻

[@BusinessScience]

Thanks! Both Modeltime & Fable are two different ecosystems for forecasting in R. Fable excels at iterative ARIMA, ETS, etc. Modeltime unlocks Machine Learning and Deep Learning. Both are great. Sky is the limit.

[@freebooterish]

Good job! Which course can I get the full version of Nostradamus?

[@BusinessScience]

Full version isn’t available. Lite version is available in Learning Labs Pro. university.business-science.io/p/learning-labs-pro

[@BusinessScience]

Also you can learn time series and shiny in the courses so you can build your own Nostradamus. Time Series university.business-science.io/p/ds4b-203-r-high-performance-time-series-forecasting Shiny Developer university.business-science.io/p/expert-shiny-developer-with-aws-course-ds4b-202a-r/

[@freebooterish]

@@BusinessScience Thanks, I am more interested in the time series course.

[@BusinessScience]

@@freebooterish it’s a good course. Every strategy used in the app comes from the course. Ensembles, machine learning, visualization. university.business-science.io/p/ds4b-203-r-high-performance-time-series-forecasting

[@atinderbharaj]

Hey Matt! How do you add external regressors ? I added them to the recipe and they definitely appear in the recipe object, but it produces an error. was hoping you can make a video regarding this. thanks!

[@BusinessScience]

I did better! I made a whole course that shows you exactly how to add external regressors when forecasting and scale up to thousands of time series. It's called High-Performance Time Series Forecasting. Learn more: university.business-science.io/p/ds4b-203-r-high-performance-time-series-forecasting/

[@alanjiang2930]

hi Matt, thanks for sharing. Really Really helps. Just wonder for the dataset you showed in the video, approximately how long on average it takes to train one submodel, lets say prophet. my R was stuck there. did you pre-load the trained submodels for the demo? thanks!

[@BusinessScience]

Training times will vary based on what model you use and your dataset. For example, Auto ARIMA can take hours to train on a single time series with multiple seasonalities and high-frequency. Xgboost can train on the same series in milliseconds.

[@Jan-gr5wl]

Hi Matt, first of all, thanks for an amazing job with modeltime! I have a question regarding the processing with several IDs, as you show in this lab. If we consider one method, is every model fitted to every time series each step by step, or is one single model fitted to all id-timeseries at once? This question also combines with the maths behind the accuracy: is it calculated as a mean of all single prediction accuracies over all IDs? This questions came up do to the fact, that the mase for fitting all my timeseries (which each have an ID to distinct) seperately with the same models provides much lower accuracy in average, which was surprising to me. Maybe it's not the best place for such a long comment, but maybe you can answer it easily or link to the code in github that can answer my questions. Thank you!

[@BusinessScience]

You are right regarding accuracy metrics. They tend to be dominated by the largest errors so it’s best to review the time series individually. We can do this and I believe it’s done in the lab.

[@Jan-gr5wl]

@@BusinessScience Thanks for your reply, you're right, the comment was a bit fast, I found the lab in which you explain the review by time series afterwards. It works well, thank you!

[@rainbowdu509]

Do you have this course available on python?

[@BusinessScience]

No it’s only in R.

[@brittnyfreeman3650]

It’s available now if you’re still interested.

[@benxneo]

Do you perhaps share the code for your app?

[@BusinessScience]

You can get the lite version of Nostradamus when you join LL PRO university.business-science.io/p/learning-labs-pro