Retail analytics · 2026
Demand Forecasting
A time-series forecasting service that cut planning error by 21% across thousands of SKUs.
Demand Forecasting
- Role
- Lead data scientist
- Timeline
- 5 months
- Team
- 2 data scientists, 1 data engineer, 1 product manager
- Focus
- Python, LightGBM, Airflow
The problem
Planners relied on spreadsheets and moving averages to forecast demand. Errors were high on promotions and slow-moving items, which led to stock-outs on some products and overstock on others.
My role
I owned the modelling approach, the evaluation framework and the production pipeline. I worked with a data engineer on features and with the product manager to agree on the metrics that mattered to planners.
Key decisions
- Start with a strong baseline. A seasonal naive model gave us a fair benchmark and kept every later gain honest.
- Gradient boosting over deep models. LightGBM with lag, calendar and promotion features was accurate, fast to train and easy to explain.
- Backtest like production. Rolling-origin evaluation by SKU group replaced a single train/test split.
Outcome
Forecast error fell by 21% against the previous process, and the weekly pipeline now runs unattended on Airflow with drift alerts. Next time I would involve planners earlier in feature design.