Our AI Demand Model Outperformed LS Central's Built-In Replenishment by 43%

LS Central's built-in replenishment engine is competent — min/max levels, reorder points, vendor lead time buffers. The problem is that it is static. It reacts to stock levels, not to demand signals. It cannot know that a competitor's promotion next week will spike a category by 40%, or that the upcoming Ramadan period will shift purchase patterns in ways that last year's data can help predict.

We built a model that knows those things.

The Training Data

18 months of LS Central transaction data from a 9-outlet supermarket group: 22 million line items, 8,400 active SKUs, transaction timestamps granular to the second. Augmented with local weather data, a public holiday and cultural calendar, the LS promotional calendar, and competitor promotional activity.

The Model

LightGBM with a multi-output structure — one output per item/store pair, forecasting daily demand over a 21-day horizon. Key feature engineering: lag features at 7, 14, and 28 days; rolling averages weighted toward recency; event proximity features; promotion state indicators; weather interaction terms for relevant item categories.

How It Connects to LS Central

The model runs nightly as an Azure ML batch job. Output is a forecast table — item, store, date, predicted demand, confidence interval. We wrote an AL extension that reads this table and converts it to LS Central purchase suggestions, respecting vendor minimums and lead times. The buying team sees AI suggestions alongside LS's standard suggestions, with the AI forecast as visible context.

Results

• Stockout frequency: reduced 43% vs. LS min/max baseline
• Overstock days: reduced 28%
• Forecast accuracy (MAPE) at 7-day horizon: 12.4% vs. 21.7% for min/max equivalent
• Buyer acceptance rate of AI suggestions: 78%
• Manual reorder adjustments per week: reduced 64%