Business Context

NovaRetail launches 200-300 new consumer products each quarter across e-commerce and stores. The merchandising team needs a model to predict first-8-week unit sales for new launches so they can set initial inventory, pricing, and marketing budgets before launch.

Dataset

Historical data contains prior product launches from the last 4 years. Each row represents a product launch in a specific market-channel combination.

• Size: 48K launch records, 48 features
• Target: Continuous — units sold in the first 8 weeks after launch
• Missing data: 12% missing in marketing plan fields, 7% missing in external trend features, and sparse history for truly new brands
• Skew: Strong right tail; top 5% of launches account for ~28% of unit volume

Success Criteria

A good solution should achieve MAPE < 22% on mature categories and weighted MAE < 180 units overall, while producing calibrated forecasts that are usable for inventory planning. The model should also identify the main drivers of predicted demand for planners.

Constraints

• Predictions are generated in batch once per week for upcoming launches
• Forecasts must be explainable enough for category managers to challenge assumptions
• The system must handle cold-start launches with limited historical brand data
• Retraining budget is moderate; a heavy deep learning stack is not required

Deliverables

1. Build a regression pipeline to predict first-8-week launch demand.
2. Explain feature engineering choices for cold-start and sparse-history products.
3. Define a leakage-safe train/validation/test strategy.
4. Report evaluation metrics by category and launch type.
5. Recommend how the model should be deployed, monitored, and retrained in production.