Dining Brands Group Automates Demand Forecasting with SageMaker Canvas Without Data Scientists

• IT team practitioners built and operated ML models using Amazon SageMaker Canvas without data scientists
• Achieved 90% forecast accuracy for three key fresh meat items, reducing data collection and forecasting time to approximately 30 minutes
• Automated periodic forecasting and integration with the ordering system (SRM) using Amazon EventBridge Scheduler and Lambda
• Separated model development and periodic operation pipelines to ensure version control and incident response efficiency

Technical Architecture and Configuration

• Configured an automation pipeline based on Amazon RDS, S3, ECR, Lambda, and EventBridge Scheduler
• Separated forecast result delivery paths: distinguished between SRM delivery for ordering and S3 storage for analysis
• Deployed and deleted the SageMaker real-time inference endpoint only at forecast execution times to reduce costs

Modeling and Data Strategy

• Operated three numerical forecasting models with different input variables depending on normal conditions, marketing, and new menu situations
• Improved training data quality by registering promotion and store event information as structured data in internal systems
• Compared model performance centered on RMSE and MAE metrics and analyzed variable importance using Column impact