Travel retail forecasting system

An airport retailer operating 50 duty-free stores across Europe and Asia was struggling with poor demand forecasts. Their existing approach was primarily based on manual adjustments and gut feel, with no systematic seasonality modeling or ability to handle promotional events.

This led to two major problems: frequent stockouts of popular items during peak travel seasons, and excess inventory of slow-moving products that tied up working capital. The category managers spent most of their time firefighting rather than strategic planning.

The company needed an automated forecasting system that could handle 500+ SKUs, account for complex seasonality patterns (tourist seasons, holidays, flight schedules), and provide reliable forecasts with minimal manual intervention.

I built an ensemble forecasting pipeline that combined multiple approaches to handle different product behaviors. For products with clear seasonal patterns, I used Facebook Prophet with custom seasonality features. For intermittent demand products (common in duty-free), I implemented Croston's method. The system also included custom business rules for promotional periods and special events.

The pipeline was built in Python and deployed with Airflow for automated weekly retraining. As new sales data came in, the models automatically retrained and generated updated forecasts. I also built a Streamlit dashboard that gave category managers visibility into forecasts, confidence intervals, and model performance metrics.

The key innovation was the feature engineering: I incorporated external signals like flight schedules, tourist season calendars, and holiday patterns from multiple countries, which significantly improved accuracy for this specific retail context.

Implementation took 8 weeks. Week 1-2 was discovery and data audit, where I identified data quality issues and gathered domain knowledge from category managers. Week 3-6 focused on model development and validation, testing different approaches on historical data. Week 7-8 was deployment and training, where I rolled out the system and trained the team.

The biggest challenge was handling the intermittent demand products (about 30% of SKUs), which required a different modeling approach than standard time series methods. I also had to build robust data quality checks, as the point-of-sale data occasionally had errors that could corrupt the forecasts.