Scaling forecast computation is a hard problem, often costs developers in productivity. Even established organizations spend a lot of time trying address this, but often the outcomes are not easy to use and manage[1][2].
We spent the last few years building a demand planning product for CPG and retail in which we had to forecast thousands of time-series. Computing such huge number forecasts in a reasonable time was a challenge. Facebook’s Prophet was one of our work-horse algorithms. Although Prophet made many things easy, it was still slow for a large dataset.
Speed was critical for quick experimentation. To improve the speed, we built a small library to run compute on AWS Lambda. Encouraged by the early results, we built an in-house service HyperProphet which we are sharing with you today.
The key features of HyperProphet are: - Fast: compute thousands of forecasts in minutes - Simple: switch from Prophet to HyperProphet by changing a couple of lines of code - Productive: experiment with forecasts in your notebooks without worrying about infrastructure
HyperProphet is in early beta, and we are looking for more people to try it and give us feedback.
We would love to see it used in diverse use cases and see how it performs. We see possible use-cases across multiple industries for capacity planning, anomaly detection, financial planning, among others.
If you have any thoughts, we’d love to hear them in the comments!
[1]: https://www.slideshare.net/MahanHosseinzadeh/prophet-at-scale-using-prophet-at-scale-to-tune-and-forecast-time-series-at-spotify
[2]: https://medium.com/walmartglobaltech/scaling-machine-learning-algorithms-fbprophet-xgboost-with-pyspark-on-w-mlp-405fadca1c19
