Six warehouse silos each sell a product whose daily demand follows a hidden seasonal wave plus random noise — the AI never sees the true wave, only past demand. It maintains a running forecast with exponential smoothing, weighting the newest observation by α and the previous forecast by (1−α), then sets a dynamic reorder point that scales with lead time and the demand it expects during that lead time. When a silo's stock drops below its reorder point, a truck is dispatched from the depot carrying an order-up-to quantity; it arrives — and the stock rises — only after the full lead time has elapsed, exactly as real freight would.
forecast(t) = α·demand(t−1) + (1−α)·forecast(t−1)
reorderPoint = forecast · leadTime · 1.15 + safetyStock(volatility)
orderQty = forecast · leadTime · 2 − currentStock
- Reorder policy — toggle between the adaptive AI forecast (reorder point tracks the seasonal wave) and a naive fixed threshold that ignores seasonality entirely, the way many legacy inventory systems still work.
- Demand volatility — how much random noise is layered onto each day's demand; higher volatility makes a fixed threshold fail more often.
- Seasonality strength — amplitude of the underlying demand wave each silo follows (out of phase with its neighbours), the pattern the AI is trying to learn.
- Supplier lead time — days between placing an order and the truck arriving; longer lead times force earlier, larger reorders.
- Forecast smoothing α — how much weight the newest day's demand gets versus history; too high overreacts to noise, too low reacts too slowly to real seasonal shifts.
Real-world relevance: this is the predictive-analytics half of AI supply-chain optimization — accurate demand forecasting keeps inventory holding costs down while cutting stockouts, the same trade-off retailers like Walmart tune for at enormous scale.