I recently chose inventory demand forecasting as my capstone project in a 10 month Data Science program I completed (at Flatiron School). I was excited to explore this topic, as I felt it could apply to a few different areas of career interest that I have. However, as I got to working on my project in applying Facebook prophet, I couldn’t help but feeling like it was way too easy.
I couldn’t help but wonder if I had done something wrong by choosing a package that basically did the work for me.
Had I in some way cheated? Would this project be displaying me more as an analyst than a data scientist?
Luckily my fears were quelled when my instructor assured me that data scientists’ main purpose is to solve business problems, which I successfully did. And that we use different machine learning tools to solve those problems. So if one tool happens to efficiently solve that problem, that’s a good thing – it doesn’t have to be complex. I essentially saved time (and in a job, this would equate to time and money). Great job, Facebook.
Upon further research, I discovered that Facebook had designed this package so that an analyst could do time series forecasting. In addition to data scientists. Time series modeling can be a specific area of specialty in the industry, so not every data scientist will already be familiar with it.
Facebook made it so that data scientists and analysts alike could produce time series models at the same level as an experienced time series modeler, efficiently and at scale. And I’m here to affirm (not that they require my affirmation) that they have indeed succeeded.
They call this their analyst-in-the-loop approach to forecasting at scale, which they say makes the best use of human and automated tasks.
My baseline model with no hyperparameter tuning was able to accurately forecast the test data with just 8.65% MAPE (mean absolute percentage error). Then, when tuning just two of the hyperparameters (changepoint and seasonality prior scales) by performing a grid search as instructed per Prophet’s documentation, I was able to improve the model to just 6.47% MAPE. This is considered excellent in product (inventory) demand forecasting.
Their documentation is one of the best I’ve seen; offering a well organized tutorial on most of the model uses for both R and Python simultaneously. For a deeper dive, they even include a link to their white paper, which explains the calculations behind the model, and displays even more potential ways to use it.
Overall it seems Facebook has successfully created an open source time series modeling package. One that makes it so data scientists new and seasoned, as well as analysts who code in Python and/or R, can now model and forecast time series data – both efficiently and at scale.
Solving business problems efficiently is always a good thing.
Have you tried Facebook Prophet‘s time series modeling package yet? What did you think about it? I look forward to hearing your responses in the comments.
Thank you for reading! I hope you enjoyed.
References:
- Prophet Docs: https://facebook.github.io/prophet/docs/quick_start.html
- Prophet White Paper: https://peerj.com/preprints/3190.pdf
