The source files are downloaded from Freshdesk once a month and saved in an S3 bucket.
Snowflake objects:
- External table for the source data
- Materialize view for daily tickets count
- Views for training, testing and forecast data
- Timeseries model to predict number of tickets for the next 30 days by customer support queues
- Task to re-train the model monthly
- SQL stored procedure to extract prediction
- Email integration
- Python stored procedure to send nice formatted email with the prediction
I did some research to add external features like public holidays but it does not impact the model accuracy. The overall results are not very impressive.
Some results:
Customer Service Queue 1
Customer Service Queuep 2
All other queues
Feature Importance
Metrics
