10.1 Forecast Horizons, Hierarchy & the Aggregation Principle
Key Takeaways
- Aggregate forecasts are always more accurate than the sum of their detailed components, because independent random errors partially cancel — the forecasting counterpart of inventory risk pooling.
- Forecast horizon drives method: short-horizon operational forecasts use statistical time series, while long-horizon strategic forecasts rely on causal models and judgment.
- Top-down forecasting builds at the aggregate level and allocates down; bottom-up builds at SKU level and sums; middle-out forecasts at the product-family level and works in both directions.
- ABC/XYZ segmentation splits items by value (ABC) and by demand variability (XYZ), and the AX-to-CZ grid determines which forecasting method and service level each item deserves.
- Intermittent or lumpy demand breaks conventional exponential smoothing; Croston's method forecasts demand size and demand interval separately.
Forecast Horizons, Hierarchy & the Aggregation Principle
Before choosing a formula, a planner must decide what to forecast, at what level, and over what horizon. Getting those structural choices wrong makes the arithmetic irrelevant, and ISM tests them directly because they are where supply management adds judgment rather than computation.
Forecast Horizon Determines Method and Purpose
| Horizon | Typical span | Time bucket | Primary method | What it drives |
|---|---|---|---|---|
| Very short (demand sensing) | 0–4 weeks | Daily / weekly | Downstream signals, point-of-sale, order book | Deployment, allocation, expediting |
| Short (operational) | 1–3 months | Weekly / monthly | Time series — smoothing, moving average | Replenishment, master production schedule, supplier releases |
| Medium (tactical) | 3–18 months | Monthly / quarterly | Time series plus causal, S&OP consensus | Capacity, workforce, contract volumes, inventory build |
| Long (strategic) | 18 months – 5 years | Quarterly / annual | Causal, econometric, judgmental, scenario | Network design, capital investment, long-term agreements, make-or-buy |
Exam anchor: statistical time-series methods lose validity as the horizon lengthens, because they assume the underlying pattern persists. A scenario asking for a five-year demand view to justify a distribution centre investment wants causal and scenario methods, not exponential smoothing.
The Aggregation Principle — Why Level Matters More Than Method
This is the single most useful and least intuitive idea in forecasting:
A forecast at an aggregate level is always more accurate, in percentage terms, than the sum of the individual forecasts beneath it.
The reason is statistical. Individual item errors are partly independent, so when items are summed, over-forecasts on some offset under-forecasts on others. If $n$ items have independent demand with equal standard deviation $\sigma$, the standard deviation of the aggregate is:
while the mean grows by $n$. Relative variability — the coefficient of variation — therefore falls by a factor of $\sqrt{n}$.
Worked illustration. Ten colour variants each average 500 units per month with a standard deviation of 150 units (coefficient of variation 30%).
- Aggregate mean = $10 \times 500 = 5{,}000$ units.
- Aggregate standard deviation = $150 \times \sqrt{10} = 474$ units.
- Aggregate coefficient of variation = $474 / 5{,}000 = \mathbf{9.5%}$, versus 30% at the item level.
The practical consequences run through the whole exam:
- Forecast at the highest level the decision permits. A capacity decision needs the family total, not the SKU split.
- Postponement exploits this directly. Holding generic inventory and configuring late means you only need the accurate aggregate forecast, not the inaccurate variant-level one.
- Risk pooling in inventory is the same mathematics. Centralizing stock across regions reduces total safety stock by roughly the square root of the number of locations consolidated.
- Never judge SKU-level accuracy by aggregate-level targets. A 5% error at family level and a 35% error at SKU level can describe the same forecast.
Forecast Hierarchy: Top-Down, Bottom-Up, Middle-Out
| Approach | How it works | Strengths | Weaknesses |
|---|---|---|---|
| Top-down | Forecast the aggregate (category or total business), then allocate down using historical proportions or planned mix | Accurate at the top; fast; good for financial planning | Allocation ratios go stale; poor at capturing item-level events such as a single SKU promotion |
| Bottom-up | Forecast each SKU-location, then sum upward | Captures item-specific knowledge, promotions, and new items | Noisy; effort scales with item count; aggregate error can compound if bias is systematic |
| Middle-out | Forecast at the product-family or brand level, then reconcile upward and downward | Balances stability and detail; usually the practical S&OP choice | Requires disciplined reconciliation logic |
Whichever direction is used, the levels must reconcile — the sum of the SKU forecasts must equal the family forecast. Unreconciled hierarchies produce the classic failure where sales, operations, and finance each plan to a different number, which is precisely what S&OP exists to prevent.
ABC/XYZ Segmentation — Matching Effort to Item
Not every item deserves the same forecasting investment. Two axes:
- ABC classifies by value or importance (annual usage value): A items are the high-value minority, C items the low-value majority.
- XYZ classifies by demand variability, usually the coefficient of variation: X is stable and predictable, Y is variable with a discernible pattern such as seasonality, Z is erratic or intermittent.
| Class | Character | Forecasting approach | Inventory posture |
|---|---|---|---|
| AX | High value, stable demand | Statistical time series with tight review; highest planner attention | Low safety stock, high service level, frequent replenishment |
| AY | High value, seasonal or variable | Trend and seasonal models; consensus overlay | Moderate safety stock, seasonal build plans |
| AZ | High value, erratic | Judgmental and collaborative; customer-level intelligence; possible make-to-order | Buffer capacity rather than buffer inventory; consider consignment or VMI |
| BX / BY | Moderate value | Automated statistical models with exception management | Standard reorder-point policy |
| CX / CY | Low value, predictable | Simple automated rules, min-max, two-bin | Larger order quantities, less frequent review |
| CZ | Low value, erratic | Do not forecast individually; stock to a simple rule or buy to order | Two-bin, kanban, or supplier-managed |
Exam anchor: the correct answer for a CZ item is almost never "build a better statistical model." It is to simplify the replenishment rule, push responsibility to the supplier, or hold a modest buffer, because the forecasting effort costs more than the error it removes.
Intermittent and Lumpy Demand
Demand for spare parts, capital equipment, and slow-moving industrial items arrives in sporadic bursts with many zero periods. Conventional exponential smoothing fails here: it averages the zeros into the forecast, producing a small non-zero number every period that is never the right answer.
Croston's method solves this by decomposing the series into two separate exponentially smoothed components:
- The size of demand when it occurs.
- The interval between demand occurrences.
The forecast per period is then the smoothed demand size divided by the smoothed interval, and both components update only in periods where demand actually occurs. Variants such as the Syntetos-Boylan approximation correct a known bias in the original formulation.
Demand pattern classification guides the choice:
| Pattern | Average demand interval | Demand size variability | Method |
|---|---|---|---|
| Smooth | Frequent | Low | Standard exponential smoothing |
| Erratic | Frequent | High | Smoothing with wider safety stock; investigate causes |
| Intermittent | Infrequent | Low | Croston's method |
| Lumpy | Infrequent | High | Croston's or judgmental; often best served by make-to-order or a service-level agreement rather than stock |
Choosing the Time Bucket
The bucket must match the decision. Forecasting in daily buckets to drive a supplier with an eight-week lead time creates noise without information. Forecasting in quarterly buckets to drive weekly warehouse replenishment hides the variation that safety stock exists to cover. A workable default: the bucket should be no finer than the shortest decision cycle it drives, and no coarser than the replenishment lead time it must cover. Rolling horizons — where the forecast is refreshed each period and extends the same total distance forward — keep the planning window constant as time passes and are the standard S&OP practice.
Ten colour variants of a product each average 500 units per month with a standard deviation of 150 units. What happens to relative forecast variability when the ten variants are forecast as a single aggregate, and what practical technique exploits this?
A planner is trying to improve the statistical forecast for a low-value spare part that shows demand in only 4 of the last 24 months, with quantities ranging from 1 to 40 units. What is the appropriate response?
An executive team asks the supply chain group for a five-year regional demand view to support a proposed distribution centre investment. Which forecasting approach is appropriate?