Hyperlocal Forecasting at Scale: The Swiggy Forecasting platform Introduction Time series forecasting on the hyperlocal level The Swiggy Forecasting platform (FP) Event Handling The End-to-End-Pipeline Tenets for the pipeline design Implementation Conclusion


Figure 1: Actuals and forecasted values of hourly demand for Indira Nagar and Chennai East zones. Here, the forecast horizon is 7 days with a forecasting delay of seven days, which suggests on the Nth we’re forecasting from N+8 to N+14th days.
Figure 2: Event Impact model as a post-processing step
Figure 3: Results from Event Impact Modeling, and the y-axis is wMAPE. Here 0,0 is for the dates of no-festivals, and no-events; 1,0 for the dates of festivals, and no-events; 0,1 for the dates of no-festivals, and events
Figure 4: The Forecasting platform’s end-to-end pipeline
  1. Run AutoML, where for every algorithm one of the best hyper-parameters are identified by evaluating on predefined backtest windows.
  2. Train a particular algorithm and construct a model by specifying model-specific parameters in addition to pre/post-processing steps
  3. Construct Composite models like tournaments and ensembles over base models
Figure 5: Actors and component view
Figure 6: XGBoost forecasting API
Figure 7: Hyper-parameter tuning API for XGBoost
Figure 8: The training stage User Interface to establish the forecasting pipeline, here tags are used as a mechanism to implement governance and manage the life cycle


What are your thoughts on this topic?
Let us know in the comments below.


0 0 votes
Article Rating
Newest Most Voted
Inline Feedbacks
View all comments

Share this article

Recent posts

Would love your thoughts, please comment.x