The context
Capsule drops create demand peaks that last two to three weeks. Peaks that short are hard for a traditional supply chain to follow. Marketing data is the earliest signal a brand has.
A hybrid model
- ARIMAX, a time-series model, with ad spend lagged by seven days as an external input.
- LightGBM, a gradient-boosting model, using spend from seven days earlier, clicks from three days earlier, Google Trends interest in the brand and a weather index.
Five steps to production
- Collect ad spend and event data every day.
- Engineer features such as lags and rolling averages.
- Split data into training and test sets with rolling-origin validation.
- Back-test forecast error (MAPE) against sell-through.
- Deploy on Vertex AI and trigger purchase-order suggestions automatically.
How to report it
- Actual against forecast demand.
- Feature importance with SHAP values.
- A heatmap of forecast error by product class.
Notes
First published on bespokemarketing.asia and updated for this site. Figures are as cited in the original article; tool defaults are examples unless stated.