What Used to Happen Has Disappeared

SSG.COM predicts product-level shipment volumes by center and store to use in automatic ordering and recommended ordering. For products with sufficient data, it forecasts using ML models such as XGBoost, LGBM, Ridge, Lasso, while for products with insufficient data, it uses moving average, exponential smoothing, and Croston methods.

However, when market changes, promotional events, and operational issues overlap, forecasts can easily become unstable. In the first half of 2025, a sharp spike in stockout rate occurred at a specific center, and an analysis of operational share and stockout rate by order type showed that automatic ordering managed more than 2x as many products as recommended ordering while maintaining a lower stockout rate. Based on this, the direction was set to gradually convert recommended-ordering targets to automatic ordering.

An examination of conversion feasibility found that for products under recommended ordering, the automatic-ordering forecast values had errors versus actual shipment volumes falling within ±0.2 for over 80% of cases. In other words, it was judged that replacing recommended-order quantities with automatic-ordering forecast values was unlikely to worsen stockouts further.

For sharp demand fluctuations during promotional periods, TFT (Temporal Fusion Transformer) was introduced. While existing models were slow to react during periods of sharp shipment increases, TFT better captured the promotion start-middle-end flow by reflecting temporal order and time-point importance, and forecast accuracy during promotional periods improved after its introduction.

Quantile Forecasting was also applied alongside this, using different forecast quantiles depending on the situation.

• Stockout-prevention focus: uses upper-quantile forecast values
• Inventory-optimization focus: uses lower-quantile forecast values

This approach allowed the system to select higher forecast values based on internal signal conditions, enabling earlier response to stockout risk. It was especially effective in situations where the risk of underforecasting increased, such as after Center A's capacity was transferred to Center B.

Results were also confirmed. The share of automatic ordering expanded from 57% in February to 77% in August, and the RMSE of the finally selected model dropped from around 11% in January to about 3% in mid-July. Products converted from recommended ordering to automatic ordering showed a 16.1% improvement in stockout rate compared to products that remained on recommended ordering.

However, an issue remains where order quantities set at high quantiles right after a promotion ends can arrive late, leading to excess inventory. The effort concludes with a commitment to continue refining forecast values and ordering criteria in line with changes in center operating conditions.