China AI Forecasts: Right Total, Wrong Stores

A demand forecast can get the chain-wide total exactly right while leaving one store short and another with unsold stock. When evaluating a Chinese retail AI system, ask whether its reported forecast score uses the same product, location and time interval as the replenishment decision.

Start with the decision the buyer can actually make. Stock that is available in one branch may not reach another before customers need it. Adding the two branches together removes that constraint from the score.

Keep the store rows visible

Consider an illustrative one-day exercise for one product at two stores. Assume no opening stock, deliveries equal to the forecast quantities, no transfers or replenishment during the day, and no substitution or backorders. Demand is fully known in this synthetic exercise.

Store	Forecast and delivered units	Demand	Units served	Unserved demand	Stock left
A	90	10	10	0	80
B	10	90	10	80	0
Total	100	100	20	80	80

The total forecast error is zero. Yet only 20 of the 100 requested units can be served from the stock at the correct store. Eighty units remain at A while 80 units of demand go unserved at B.

Define signed forecast error as forecast minus demand. A's error is +80 and B's is -80. Their sum is zero, but the sum of their absolute errors is 160 units. That 160 measures mismatch across the two forecast rows. It is not 160 lost sales. Under these assumptions, unserved demand is 80 units. Keeping those quantities distinct prevents an error metric from turning into an exaggerated commercial claim.

Compare the alternative at the same decision level

Now suppose a baseline allocates 50 units to each store. It also predicts 100 units in total. Under the same demand and operating assumptions, A serves ten and retains 40; B serves 50 and leaves 40 units of demand unserved. The baseline serves 60 units altogether.

Both plans have a perfect total forecast. Their store-level consequences differ. Ask the supplier to compare the AI proposal and the current planning process on the same store-product-day rows, with the same information cutoff and operating constraints. Preserve the original forecast rather than replacing it with an update made after demand became visible. The forecast backtest guide covers that timing boundary.

Record what can actually move

If the retailer can transfer stock during the day, add the transfer departure and arrival times, quantities, capacity and cost to the exercise. Count a transfer as serving demand only when the stock reaches the required location in time under the agreed service rule. An available van does not establish that every shortfall can be recovered.

Apply the same discipline to time aggregation. Surplus stock arriving tomorrow cannot satisfy a requirement that had to be met today. Choose the reporting interval to match the decision and delivery window, and retain the finer records needed to investigate a mismatch.

The example deliberately assumes known demand. In real records, sales during a stockout may omit customers who could not buy. Mark stock availability and any independently recorded unmet requests. If lost demand is estimated, show the method and uncertainty separately from observed sales. Do not calculate a precise service rate from a denominator the team has not measured.

Request two reports from the pilot

Keep a total-level report for purchasing and capacity discussions, and a decision-level report for allocation. The latter should retain forecast, baseline, stock available before the deadline, observed sales, evidence of unmet demand, transfers and remaining inventory for each relevant location and period.

NIST AI RMF 1.0 (2023), MEASURE 2.1 calls for documenting test sets, metrics and evaluation tools. For this retail exercise, document the aggregation level beside every score. The buyer should be able to trace a good headline result down to the stock available where and when the customer needs it.