5 Benefits of Modeltime for Time Series Forecasting

I built modeltime to make time series forecasting more efficient and reproducible. Now you get the benefits. Here are 5 reasons why modeltime for time series forecasting!

Transcript

1. Why modeltime?
2. Why I made modeltime I built modeltime for me to: • Increase my productivity • Decrease my code • Improve my ability to teach you how to forecast It’s really good. I want you to have it. Modeltime: https://github.com/business-science/modeltime
3. First, my roadblocks. I’ve tried a lot of time series software... My forecasting adventures in R & Python Individual Modeling Approaches (Roadblocks 1 & 2): • Forecast (ARIMA, ETS, TBATS) • Prophet • BSTS • GARCH • Deep Learning (TensorFlow, PyTorch, mxnet/gluon) Fable/Tsibble Forecasting Ecosystem • Great, but see Roadbock 3 Tidymodels Machine Learning • Excellent, but see Roadblock 4 My roadblocks 1. Too many data structures (costs 2-lines of code for every conversion): ◦ Data frame ◦ Tsibble ◦ Xts ◦ Zoo ◦ Ts ◦ Matrix 2. Inconsistent approaches (leads to switching costs & productivity loss). 3. Focus is either on ARIMA or Machine Learning, not both (and why not Deep Learning?) 4. Tidymodels Machine Learning is on the right track, but didn’t have a forecasting toolchain or forecasting models
4. Typing 1000 lines of code to make a multi-model forecast
5. Consolidating my multi-model forecast into 200-lines of code
6. 5 Benefits • • • • •
7. Leverages Tidymodels Gives me access to: • Parsnip - Machine Learning Models • Recipes - Preprocessing • Workflows - Organizes models & recipes • Tune & Dials - Hyper parameter tuning • Yardstick - Accuracy measures Source: https://www.tidymodels.org/ 1 Benefit
8. Leverages Tidymodels Source: https://www.tidymodels.org/find/parsnip/ Parsnip Ecosystem: Over 20 Machine Learning Algorithms and growing 1 Benefit
9. Integrates Forecast & Prophet Forecast Prophet Time series models I’ve currently included as of Modeltime 0.0.2: • ARIMA, Exponential Smoothing, & Seasonal Decomposition • Prophet More models coming soon 2 Benefit
10. Simple Forecasting Workflow MODELTIMEWorkflow Create Modeltime Table modeltime_table() Calibrate modeltime_calibrate() Forecast Test Set modeltime_forecast() plot_modeltime_forecast() Test Accuracy modeltime_accuracy() table_modeltime_accuracy() Refit modeltime_refit() Forecast Future modeltime_forecast() plot_modeltime_forecast() 3 Benefit
11. Fits nicely inside a Time Series DS workflow 4 Benefit Hadley & Garrett’s essential data science workflow from R for Data Science
12. Fits nicely inside a Time Series workflow Timetk: https://business-science.github.io/timetk/ 4 Benefit
13. Fits nicely inside a Time Series workflow Modeltime: https://github.com/business-science/modeltime Timetk: https://business-science.github.io/timetk/ 4 Benefit
14. Every output is interactive (or static) plotly ggplot 5 Benefit
15. Kick the tires Test drive modeltime Forecasting Tutorial https://www.business-science.io/code-tools/2020/06/29/introd ucing-modeltime.html
16. What’s coming • • •
17. Modeltime Advancement Roadmap ✅ Phase 1 - Initial Release Phase 2 - More models & ensembling • Fill time series model gaps • Tidymodels stacking framework • Deep Learning Comment here if you want a model / algorithm / package: https://github.com/business-science/modeltime/issues/5
18. Advanced Time Series Course (Coming Soon) Prerequisite: R for Business Analysis (DS4B 101-R) 3 Part - Forecasting Training 1. Exploration & Feature Analysis ◦ Data Processing ◦ Visualization ◦ Feature Engineering 2. Time Series Modeling ◦ Automated Algorithms ◦ Boosting & Scaling ◦ Machine Learning Algorithms & Tuning ◦ Ensembling 3. Deep Learning ◦ GluonTS ◦ Python & Reticulate Learn Forecasting Most advanced technology to deliver high-performance forecasts Join the Course Waitlist