7. Should We Reconcile Forecasts to Align Supply Chains?
Teaching rewrite — the Objective step, across teams: one shared forecast, or aligned but different forecasts? Original wording; numbers are illustrative.
Introduction
Supply chains need teams to act in concert: if marketing launches a product in Q2, supply must have stock by the end of Q1. A popular answer is the one number forecast — a single reconciled forecast everyone uses. An alternative is the one number mindset — shared data, assumptions, and events, but each team keeps the forecast that best fits its decisions, within tolerance of the others.
This lesson covers:
- Why different roles need different granularities and horizons
- The one number forecast and its three challenges
- Why the forecast of the sum ≠ sum of the forecasts (spot sales, life cycles)
- A worked top-down vs bottom-up comparison
- The one number mindset in practice
7.1 Different roles, different forecasts
| Role | Decision | Granularity | Horizon |
|---|---|---|---|
| Logistics manager | Which stores to ship to | SKU × store × day | Next days |
| Plant manager | What to produce | SKU × global × week | Next weeks |
| Marketing strategist | Which products to renew or launch | Brand × channel × quarter | 6–9 months |
| Finance manager | Cash-flow plan | Revenue × business unit × quarter | Next quarters |
Pattern: short-term = granular, long-term = aggregated. How do you keep all these aligned?
7.2 The one number forecast
One reconciled forecast, shared by everyone; different views are produced mechanically with top-down, middle-out, and bottom-up (lesson 5).
| Challenge | Why it bites |
|---|---|
| Efficiency | Reconciling weekly forecasts for thousands of SKUs over 18 months, every month, is a heavy recurring process |
| Optimality | The best forecast at one level is not the best at another — one size can’t fit all |
| Alignment | Agreeing a number doesn’t guarantee everyone uses it. Case studies describe companies that agreed a single S&OP number while departments still ran their own private forecasts for raw materials or daily logistics |
7.3 Different levels, different optimal forecasts
The forecast of the sum is not the sum of the forecasts.
Forecast a family of products A, B, C two ways:
- Bottom-up: forecast A, B, C separately and add them.
- Direct: forecast the family total as its own time series.
The two totals differ — not just mathematically, but for business reasons:
Spot sales and stock clearances
A family can look steady while individual SKUs show one-off spikes: flash promotions, surplus clearances, one-time contracts. A model can’t predict a clearance on a SKU that hasn’t sold for months. A family-level forecast can assume “some spot sales will happen” — but no top-down rule can say which SKU they will hit.
Also: don’t feed past clearances into future forecasts as if they were normal demand, or you will replenish dying items.
In the 1990s a carmaker had too many green cars. Sales pushed them out with heavy discounts. Manufacturing, seeing green-car sales rise and unaware of the discounts, scheduled more green cars. The signal was promotion-driven clearance, not real demand.
Product life cycles
Engines forecast from history: they spot a product declining toward end-of-life but cannot foresee the new SKU that will replace it. Long-term bottom-up forecasts therefore drift low. A stable family forecast (top-down) doesn’t suffer this.
So an 18-month family forecast may be good, but translating it to SKUs is hard — you don’t know which SKUs will make up the family’s future sales.
Checkpoint: one number forecast
Checkpoint — sections 7.1 to 7.3
1 / 4Which role typically needs the most granular forecast?
Pick an answer to see if you are right.
7.4 Worked example: top-down vs bottom-up
A supermarket sells about 10,000 fruits a month: mainly apples and pears, plus a few odd fruits. This month, besides apples and pears, it sold 500 kiwis (not usually stocked in quantity).
| Fruit | Actual demand | Bottom-up forecast | Abs. error | Top-down forecast (10,000 scaled by apple/pear shares) | Abs. error |
|---|---|---|---|---|---|
| Apples | 4,500 | 4,500 | 0 | 4,737 | 237 |
| Pears | 5,000 | 5,000 | 0 | 5,263 | 263 |
| Kiwis | 500 | 0 | 500 | 0 | 500 |
| Total | 10,000 | 9,500 | 500 | 10,000 | 0 |
| Sum of SKU errors | 500 | 1,000 |
- At the family level, top-down wins (error 0 vs 500).
- At the SKU level, bottom-up wins (500 vs 1,000).
- Scaling the apple/pear forecast up to 10,000 doesn’t help SKUs — it adds error, because the extra 500 belongs to some unpredictable “odd fruit”, not to apples and pears.
Scale factor check: 10,000 / 9,500 ≈ 1.0526; 4,500 × 1.0526 ≈ 4,737; 5,000 × 1.0526 ≈ 5,263.
Reconciling quantile forecasts is even harder: item demands are correlated (store-wide promotions lift everything; customers switch between fruits), so the distribution of the total is not the sum of the item distributions — e.g. the 95th percentile of apples + pears is not the 95th percentile of apples plus that of pears. Point (mean) forecasts, by contrast, add up cleanly. Forecasting competitions on hierarchical retail data (e.g. M5, 2020) asked for separate distributions per level for this reason.
7.5 The one number mindset
Instead of forcing every team onto one number, align everyone on one set of data, assumptions, and events — and let each team forecast at its own level.
| Rule | In practice |
|---|---|
| Same data | One reconciled source of demand history, prices, and master data. Two teams must never forecast from different histories |
| Same events | A formal channel — typically S&OP — to share price changes, campaigns, launches, competitor moves |
| Visible forecasts | Every team can see the others’ forecasts |
| Tolerance band | Small differences are natural; large ones (e.g. > 5% at global level) trigger a review |
Benefits: each team uses the model and granularity that fit its decisions, forecast quality improves, and meetings shift from arguing over numbers to discussing plans.
Worked example: S&OP vs weekly warehouse forecast
- S&OP forecast: country × month — 12,000 units for next month.
- Logistics forecast: warehouse × week, built from the same history with a different model — weeks sum to 12,480 for the month.
Gap = 480 / 12,000 = 4% → within a 5% tolerance; keep both.
If logistics had summed to 13,200 (10% over), planners would investigate — e.g. a promotion known in logistics but not discussed in S&OP — and either share the information or scale the weekly forecast toward the S&OP total (× 12,000/13,200 ≈ 0.909).
Checkpoint: one number mindset
Checkpoint — sections 7.4 and 7.5
1 / 4In the fruit example, why does scaling apples and pears up to 10,000 hurt SKU accuracy?
Pick an answer to see if you are right.
Rank: choosing an alignment approach
Alignment decisions
1 / 2Rank these situations from MOST suited to a strict one number forecast to MOST suited to a one number mindset.
OrderDrag the handle to rearrange — top should be first.
- 1Mid-size company with monthly S&OP and a few regional warehouses
- 2Small company, one warehouse, monthly decisions only
- 3Large company deciding daily store deliveries and 12-month raw-material buys
Drag items into the correct order (or use ↑↓), then check.
Check: reconciling forecasts
Should we reconcile forecasts?
1 / 6What is a one number forecast?
Pick an answer to see if you are right.
Takeaways
- Different roles need different granularities and horizons
- A one number forecast helps clarity but costs efficiency, sacrifices optimality, and doesn’t guarantee alignment
- The forecast of the sum ≠ sum of the forecasts — spot sales and life cycles show up at different levels
- Forcing SKU forecasts to match a family total can destroy SKU accuracy
- A one number mindset aligns data, assumptions, and events; teams keep fit-for-purpose forecasts within a tolerance band
Further Reading
- 5. Forecasting hierarchies
- 6. Forecasting horizon
- International Business home
- Next: 8. Forecasting metrics
- Background: N. Vandeput, Demand Forecasting Best Practices (Manning) — study outline; S. Clarke, “One-number forecasting” (Argon & Co, 2019)