Design a Demand Forecasting System for Dark Stores

Problem

Design the demand forecasting system that predicts how much stock each dark store needs for the next 24–72 hours.

Requirements

Functional:

• Forecast demand per (store, product, time_window)
• Trigger auto-replenishment orders when projected stock falls below threshold
• Support manual override by store managers
• Incorporate real-time signals (weather, local events, flash sales)

Non-functional:

• 300 stores × 5000 SKUs = 1.5M forecast cells updated hourly
• Forecast accuracy target: MAPE < 15%
• Replenishment trigger latency < 5 minutes

Discussion points

1. Offline training pipeline (feature engineering, model selection)
2. Online feature store for real-time signals
3. Forecast serving: batch precompute vs on-demand
4. Feedback loop: actuals vs forecasts for model retraining
5. Cold start for new SKUs and new stores