You are building demand forecasting for a grocery retailer with 2,400 stores and 60,000 SKUs. The output drives automated replenishment orders: too low and shelves are empty, too high and fresh produce is thrown away.
You need a forecast for every store-SKU pair for each of the next 14 days, refreshed nightly. That is 144 million store-SKU combinations, though only about 40 million are active in any given week.
The demand signal is messy. Most store-SKU pairs sell 0 to 3 units a day. Promotions cause 5-10x spikes. Weather moves ice cream and soup in opposite directions. And crucially, your history is censored: if a product was out of stock, recorded sales were zero, but true demand was not.
Build the architecture on a canvas: place the components, configure them, connect them into a data flow, and write a short reason for each one. The AI reviewer grades your design against a rubric written specifically for this problem.
A product was out of stock for 5 of the last 14 days. What does your training data say happened, and why is that dangerous?
Why is MAPE the wrong metric here, and what do you use instead?
Replenishment asks for one number per store-SKU. Which quantile do you give them for bananas versus for canned soup, and why?
Minimum 6 components · needs a wide desktop screen