DemandCast

Large-scale demand forecasting pipeline — LSTM with Temporal Attention achieving 22% lower MAPE across 50K+ products

End-to-end demand forecasting system for predicting daily SKU-level demand across 50K+ products, comparing classical and deep learning approaches.

What I built:

  • Benchmarked ARIMA, Prophet, XGBoost, and LSTM — LSTM with Temporal Attention won with 22% lower MAPE vs ARIMA baseline
  • Real-time inference API with FastAPI + Redis caching for sub-100ms predictions
  • Kafka-based streaming pipeline to ingest live sales data
  • Model monitoring layer that detects data drift and triggers automated retraining

Tech stack: Python · PyTorch · XGBoost · FastAPI · Kafka · Redis · PostgreSQL

Metric Value
MAPE improvement 22%
Products modeled 50K+
Status Ongoing