Using Transformer Models to Predict Demand at Scale in Production
summary
Predictive system uses Temporal Fusion Transformer to forecast online demand for Tchibo with a 12-week horizon, handling many short-lived SKUs by learning across similar time series. Data pipelines consolidate categorical and numeric inputs into a feature store in BigQuery, Apache Beam/Dataflow applies TensorFlow Transform preprocessing, MLflow stores models and preprocessing state, and Terraform plus GitLab deploy an event-driven microservice inference pipeline that writes predictions back into SAP. Solution runs at scale in GCP and meets the primary-warehouse service objective.
AI-written summary · about
chapters (9)
- 0:00Project introduction
- 3:13Logistics demand challenge
- 5:44Why traditional methods failed
- 7:30Temporal Fusion Transformer choice
- 9:14Cloud architecture stack
- 11:24Prediction pipeline workflow
- 13:51Microservices and model management
- 15:39Deployment and scalability
- 17:25Closing remarks
related talks
29mProbabilistic demand forecasting at scale
39mProductionising Data Science at Scale
28mUse case in demand forecasting
28mDemand Forecasting: Accuracy, Stability, Automation
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
16mTransforming Inventory Forecasting with Production ML
40mMulti-Series Forecasting at Scale with StatsForecast
30mReal-Time ML services: From ideation to production