Demand Forecast Presentation to a Skeptical COO
Northwind Supply forecasts weekly demand for 14 distribution centers. The new model has a mean absolute percentage error of 14.2% versus 22.6% for the current naive seasonal baseline, a result the data team considers a clear win. But bias is +6.3% during promotional weeks, and promotional flags are missing for three of the distribution centers. The COO wants a single forecast number for the board and is asking whether the company can cut safety stock by 15%, which would release roughly $4.2 million in working capital. Inventory carrying cost is 18% per year, and stockouts are estimated to cost $35 per unit short, though that figure is contested. The model was validated on 156 weeks of data, and the largest distribution center changed its delivery schedule eight weeks ago. The COO has already told the CFO that the model is "14% accurate."
1. How do you correct the "14% accurate" framing and present the result to the COO and board?
Read the worked answer
"14% accurate" is a category error: MAPE is an average absolute percentage error, not a probability of being right, and 14.2% is a weighted average that hides segment and period variation. I would reframe in decision terms: "The model reduces average forecast error from 22.6% to 14.2% on 156 weeks of backtest, but it is biased high by about 6% during promotions and has not been validated since the largest DC changed its delivery schedule."
For the board, one number is possible only if it is defined. For example: "Expected forecast error in non-promotional weeks is about ±14%; promotional weeks carry higher error and upward bias." I would pair the number with a range and a statement of what it excludes. A single point forecast without an error band invites the board to treat it as certainty.
I would also separate forecast accuracy from inventory policy. A better forecast does not automatically justify a 15% safety-stock cut; safety stock depends on service-level targets, lead-time variability, and stockout cost. The 14.2% error is an input to that decision, not the decision itself.
2. Should the company cut safety stock by 15%? Lay out the analysis you would do before answering.
Read the worked answer
I would model the safety-stock question explicitly rather than endorse a 15% cut. Under a simplified policy, safety stock scales with z × sqrt(lead-time variance × demand variance), so a 15% reduction changes the service level nonlinearly. Using the current error distribution, I would simulate weekly demand and compute expected stockouts and fill rates at 0%, 5%, 10%, and 15% cuts, with and without the improved forecast.
The economics: carrying cost is 18% per year on $4.2 million, about $756,000 annually, but that saving materializes only if inventory actually falls and does not merely shift. Stockout cost at $35 per unit is contested, so I would present a range - say $15 to $60 - and show the break-even service level. If the stockout estimate is uncertain, a monitored 5% pilot at two or three DCs is more defensible than a company-wide 15% cut.
I would fix the promotional bias first. A forecast that runs 6.3% high in promo weeks builds excess inventory exactly when working capital is tied up; correcting that may release more cash than a blanket safety-stock cut, without raising stockout risk.
3. The model has no promotional flags for three distribution centers and the largest DC changed its schedule eight weeks ago. How do you communicate these limitations without undermining confidence?
Read the worked answer
The missing promotional flags are a data-quality gap that directly affects the claim. I would treat those three DCs as out of scope for the promotional-bias estimate until flags are reconstructed from order history or promotion calendars, and say so explicitly. Then present three tiers: validated (11 DCs, non-promo weeks), partially validated (11 DCs, promo weeks), and not yet validated (3 DCs).
I would avoid a single blended MAPE, because it lets a good result in easy weeks mask a bad result in hard weeks. Instead, show a small table by DC and by promo versus non-promo, plus a headline range. For the board: "Average error is about 14% in normal weeks; promotional weeks are less reliable and currently biased high; three DCs need data fixes before we rely on them."
Confidence comes from precision about scope, not from hiding gaps. I would set up weekly monitoring of bias and MAPE by DC with an action trigger, and commit to revalidating the largest DC after it has eight to twelve weeks of post-change data. That gives the COO a defensible story: the improvement is real, conditional, and being actively managed.