modeltime: Time series forecasting in R with tidymodels

[@jgeise]

Great package for forecasting. Thanks, Matt!

[@BusinessScience]

Thanks so much J G!

[@88fjoe]

Great package and tutorials!

[@BusinessScience]

Thank you!!

[@siavashk100]

Thank you for this tutorial. Is there any way to adjust some parameters of auto_arima (e.g. max.p, or max.q) in this method?

[@maxirojo7829]

Hello excellent video!. At minute 8:50, how do I put the same colors on the legend labels? Thank you

[@BusinessScience]

Thanks for the kind comments. I see that the colors match but there is gray behind the color because we have a gray ribbon for the confidence intervals. To make the colors more clear you can turn off the confidence intervals. Check out the plot_modeltime_forecast() function.

[@maxirojo7829]

@@BusinessScience Thanks for the prompt response! The video is really excellent and it has been very useful to me. I am working with a time series of savings account deposits with daily data with only business days (between 21 and 22 data per month). The series has a positive trend and local lows on the first business day and local highs on the last business day (day 21 or 22). How should I calibrate seasonality inside prophet_reg? Really thank you very much

[@PA_hunter]

Awesome package. That interactive plot is also really cool! I’m not too familiar with time series…how would you handle forecasting customers for the year ahead when working with five years of data and COVID as a separate phenomenon over the last two years only?

[@BusinessScience]

Awesome glad you are loving modeltime and interested on using it to forecast with outliers present like COVID. I actually did a full training on this with 8 airlines showing how to forecast passenger load. It’s in a paid program called Learning Labs Pro. It was Learning Lab 60. university.business-science.io/p/learning-labs-pro I’ll check to see if I opened that one up.

[@BusinessScience]

I actually opened that lesson up. You can check it out here: kzbin.info/www/bejne/h5nYoHaBnbZonMk

[@PA_hunter]

@@BusinessScience awesome thanks!

[@517127]

Since modeltime extend tidymodels, can i use explainable methods used in tidymodels like vip, DALEX and DALEXtra? It would be amazing explain the forecast with modelStudio

[@BusinessScience]

Yes you can! That’s the beauty of this whole system. Anything you can do with a tidymodel you can do with modeltime. So explainable models are easy to do with vip or DALEX.

[@mohammadsaifuddin9459]

Love this :)

[@BusinessScience]

Excellent!! 😀

[@sebastiansantiago3783]

Doing the proper code, what we do grid search for the parameters?

[@BusinessScience]

We just added a new tutorial on Hyperparameter Tuning. I’d start here. It also covers parallel processing and the new tidymodels workflowsets integration. business-science.github.io/modeltime/articles/parallel-processing.html

[@sebastiansantiago3783]

@@BusinessScience hi…last question jaja I was reading about see we can use prophet with Modeltime and also is a boosted version, now Facebook deployed Neural Prophet ( I also see we have N Beats and Deep AR), is this something we as an user can include within this framework or this would be included on a newer version of Modeltime 🤩 Thanks and Regards

[@BusinessScience]

Soon!

[@jossrdgz8]

Great tutorial I am gonna register for the course! One question I usually deal with data that is reported yearly how can I define the model to use yearly step data and not monthly?

[@BusinessScience]

Yearly data that has one timestamp per year is fine. The index numeric feature will define a trend feature for the model. You can also add other features like splines to help. I teach these in my time series course.

[@wilsoncarrillo8623]

Hi...great work! can i do it with irregular time series?

[@rafaelbaptista7337]

Great tutorial. Please, how can I add regressor? I need to include a xreg to help improve the prediction.

[@BusinessScience]

You can add xregs into the data and then model the data using recipes R package for preprocessing and parsnip / workflows to create the model(s)

[@anitahaique]

Hi, Why aren't forecast intervals getting wider as we move further into the future? Thanks

[@cgonzalezcon]

Great video, thanks for the easy to follow explanations. I have been getting an error once I get to the ARIMA model: Error in `check_case_weights()`: ! 'case_weights' should be a single numeric vector of class 'hardhat_case_weights'. Do you know what it means?

[@BusinessScience]

Thanks! I’m not familiar with that error

[@patuko5304]

Hi! Please could you review the time_series_split function? There is something wrong on how the function is making the split (asses property not really taking dates into account and maybe using sample count?). If your dataset samples don't have daily registers the test date range will not be as stated. Have mapped dividing by items and got diferent test date ranges for each item. Anyway, thank you a lot for your work and teachings!

[@BusinessScience]

Can you please file a reproducible github issue at the timetk github website? This is the proper method for handling bug and feature requests.

[@Ottozone1]

On the plot_modeltime_forecast output, it is a bit difficult to see the colors on the legend. Is there parameter to change in make the color larger or standout more on the legend?

[@BusinessScience]

You can render as a ggplot object using .interactive = FALSE. Then adjust the plot to your liking. You can convert to interactive using ggplotly(). If you'd like to learn more, I have a full course on time series: university.business-science.io/p/ds4b-203-r-high-performance-time-series-forecasting/

[@jerrickren3802]

when I try to run'modeltime' Error shoed as: package or namespace load failed for ‘modeltime’: .onLoad failed in loadNamespace() for 'modeltime', details: call: check_model_doesnt_exist(model) error: Model `prophet_reg` already exists. Could U pls tell me how to resolve it?

[@emiliostar5]

Hi, does this support Neural Net based time series and Bayesian Structural Time Series?

[@BusinessScience]

Yes and yes. See modeltime.gluonts for deep learning. One of my students created bayesmodels package for Bayesian structural time series.

[@gauravmodi12]

can we do BSTS here ?

[@BusinessScience]

Yep - there is an R package called bayesmodels that includes BSTS

[@ericmai6798]

How well do these packages work with annual data rather than daily/monthly/quarterly?

[@BusinessScience]

Very good. And you can do any type of algorithms from ARIMA to xgboost.

[@PAULSHOWTO3]

how to check the MAPE and RMSE of forecast?

[@BusinessScience]

Run modeltime_accuracy() on a calibrated modeltime table

[@PAULSHOWTO3]

Using date_var: Date Error in `dplyr::group_by()`: ! Column `Date` (index) can't be a grouping variable for a tsibble. i Did you mean `index_by()`? i got this error when i terminate R session please help

[@BusinessScience]

You are using a tsibble. Try converting to tibble.

[@aparnasnayanar2621]

I am using the nbeats model for time series forecasting. Everything looks fine except am getting some rlang error : can’t find nbeats environment. When clear the state and reattach the cluster in databricks , the error goes. Can someone suggest how the error can be fixed

[@BusinessScience]

I’d start with the documentation. We have comprehensive installation and environment setup. Make sure you are using the development version business-science.github.io/modeltime.gluonts/

[@aparnasnayanar2621]

I have used the Dev version. But now the error is related to a dependency package ‘htmltools’. This is conflicting with my existing sparklyr package . I think the sparklyr needs a htmltools version higher

[@NotMuchHere]

greetings, Is there a way to see the raw data for the predictions. The variables show model information, not the raw projected data ..

[@BusinessScience]

Yes it’s output in modeltime_forecast()

[@NotMuchHere]

@@BusinessScience Thanks, that got me started. I was able to see the projections. A bit odd projections were at the beginning of the date range ... at the end of the tibble. (You have given me something interesting to play around with (figure out -- F1). Of course I changed the source data so that part is on me ... (I am going to keep trying)

[@elnarasryva6822]

Why you chose DTEDAY AND CNT ? In my data how can I choose columns

[@BusinessScience]

Good question, these are the time stamp and target value columns, respectively.

[@Lirim_K]

Never really understood the use of time series forecasting that merely replicates seasonal patterns. I could have come up with those forecasts by my self just by looking at the series previous months. So many advanced algorithms yet they do not perform better than a human.

[@hansmeiser6078]

Nice intro, but as always, no param tunings- not the whole Enchilada.

[@BusinessScience]

Just an intro. Check out the articles on our modeltime website Hyperparameter Tuning: business-science.github.io/modeltime/articles/parallel-processing.html Panel Data: business-science.github.io/modeltime/articles/modeling-panel-data.html Autoregressive Forecasting business-science.github.io/modeltime/articles/recursive-forecasting.html

[@sebastiansantiago3783]

Hi Matt, love all the material you posted and I'm enrolled on your Time Series Forecast course. Quick question, does your package works with Fable (upgrade from Forecast Package) kzbin.info/www/bejne/bKeWkHSpl82kqKc Or a totally different thing ? The reason I'm asking is because to my understanding Fable is Forecast on roids lol, and combining your package and this makes sense :)

[@BusinessScience]

Hey, awesome to hear you are in my Time Series Course. You will love it.

[@BusinessScience]

Q1: Fable - Modeltime does not currently work with Fable. We like fable for easily iterating arima and other classical models. We developed Modeltime for bringing the Tidymodels ecosystem to forecasting so we can replicate solutions to problems like the M5 Competition. In M5, XGBOOST made panel models that forecast all time series at once. It’s a really powerful idea. We also do local models. And the workflow is designed for rapidly testing and experimentation with many models.

[@BusinessScience]

Q2: Fable is the tidyverse version of forecast R package. Modeltime is tidymodels applied to forecasting. Both are insanely powerful depending on your needs. Our machine learning and deep learning ecosystem is quite extensive. And we have some more innovations coming! 😊

[@sebastiansantiago3783]

@@BusinessScience thank you for taking the time for responding which such detail…seems like I will be focusing more on learning Modeltime for now…can wait for what is coming 😱

[@djangoworldwide7925]

This is not quite a tutorial. That's crazy. As a first year stat student I kinda didn't understand any of the mL functions..

[@BusinessScience]

Understanding time series can be complex… but if you want to go deeper, I offer a very powerful course.