Choosing a forecasting method when MAPE cannot see the problem

MAPE breaks down on exactly this data.

• Division by zero. SKU 4471 has many zero months; MAPE is undefined there and, if you clip it, the metric is dominated by the small months rather than the big ones that actually drive inventory.
• It is asymmetric. Over-forecast error is capped at 100%, under-forecast error is unbounded, so MAPE rewards over-forecasting on a spike item - which is precisely what creates the write-off and the phantom replenishment.
• It hides bias. A method can show a respectable MAPE while being consistently 9% high, as SKU 2210 proves.
• It is probably in-sample. A 22% demo figure on the same history is a fit, not a forecast.

Measure instead: scaled absolute error (MASE or MAE divided by the naive-method MAE), a signed bias statistic, and forecast value added against both the naive benchmark and the current method - Mei's overrides already beat the system by 5 points, so any new tool must beat that, not just the moving average. Require rolling-origin out-of-sample backtests over at least 12 months, evaluated at the level where the decision is made: SKU-DC-week for replenishment, family-month for S&OP. Ask the vendor what loss function they optimised and to show performance on the intermittent SKUs specifically.

Treat it as two demand streams, because it is two demand streams.

The three-month moving average smears the 1,800 spike across three months, so it generates roughly 600 units of phantom demand in the two quiet months after every promotion - forecast error that becomes inventory. Split the problem:

• Baseline (the 0, 40, 25, 60 pattern): intermittent demand. Use Croston or SBA, or a simple rate-based method, and measure it with a scaled metric. Do not use a moving average on a series with this many zeros.
• Promotion (the 900 and 1,800): an event model driven by causal inputs - the two chains' feature calendars, price and display plans, and their own commitments. Get their order intentions 8-12 weeks ahead, before the buy is locked.

Process discipline: one accountable owner per stream; overrides carry a written reason code and are measured for FVA so Mei's instinct is validated or retired on evidence, not on hierarchy; and a post-event review that records actual lift against planned lift for the next event.

Where the spike is genuinely unpredictable, the answer is not a better point forecast but flexibility: short-cycle replenishment, consignment stock at the retailer DC, and a negotiated exit or return path for unsold promo packs.

Seven consecutive months in the same direction is a systematic error, not noise, and safety stock cannot fix it - extra stock only converts bias into surplus, which is what the 24% overage is.

Work the diagnostic tree in order:

1. Master data: pack-size change, a delisted customer still flagged active, duplicate item records, unit-of-measure mismatch between the forecast and the order.
2. History contamination: returns booked as demand, promotional pull-forward, sample shipments, or a large one-off order that the trend is now extrapolating forever.
3. Model: a genuine level shift the smoothing constant is chasing too slowly, or a seasonal index fitted on a distorted year.
4. Incentive: sales inflating the consensus number in S&OP.

Correct at the source first, then de-bias the forecast (refit on a cleaned history, or apply a documented bias correction until the history is fixed). Add a governance trigger: if absolute bias exceeds 5% for three consecutive months, the item goes to exception review with a named owner and a decision - this is how you stop bias surviving in a system nobody is watching.

Finally, report bias and MAPE side by side every month. A low MAPE routinely hides a large cumulative bias, and bias is the error that actually accumulates in inventory.