SKU-Level Forecasting: Approaches for Large Assortments

SKU level forecasting means producing a forecast for every stock keeping unit, often by location, and it becomes hard when an assortment grows to thousands of items. A company with 20,000 SKUs across ten locations has 200,000 item-location series to manage. No team can review each one monthly, and many of the series are noisy, intermittent or new. The answer is not to forecast everything the same way or to give up on item detail. It is to use hierarchies, segmentation and reconciliation so that the signal is found where it is strongest, the detail is filled in sensibly and human attention goes to the exceptions. This guide explains the main approaches, works through a proportional reconciliation example with checked arithmetic, covers intermittent, new and retiring items, and describes how to automate and govern the process. It is part of demand planning and inventory optimization, follows how to improve forecast accuracy and forecast value added, and leads to forecast error root cause analysis and forecast accuracy metrics. The inventory this drives is a major part of the cash described in cash flow management.

Why SKU detail is hard

Three things make item-level forecasts difficult. Volume per series is low, so random noise dominates. A product that sells ten units a week in a location swings widely, and a model fitted to ten data points per month learns the noise. Many series are intermittent, with weeks of zero sales between orders. Assortments churn, with launches, replacements and discontinuations, so history is short or broken. At the same time, replenishment decisions are made at the SKU-location level, so you cannot avoid the detail. The skill lies in borrowing strength from higher levels.

The hierarchy

Define the levels that matter: total company, division, category, family or brand, SKU, and then location or channel. Decisions are made at different levels. Budgets and capacity use the top of the hierarchy. Replenishment uses the bottom. The forecasts at all levels must be consistent, so that the SKU forecasts add up to the family forecast and the family forecasts add up to the category and total. Without this, the finance view and the operations view diverge.

Three ways to use the hierarchy

Top-down. Forecast at a high level, where the data is steadier, and allocate down to SKUs using historical proportions. It is stable and efficient, and it is blind to SKU-level changes in trend or mix. It suits mature, stable families.

Bottom-up. Forecast each SKU separately and sum upward. It captures SKU-level patterns and is sensitive to noise, and the aggregate may be poor because noise adds up in the wrong places. It suits high-volume SKUs with distinct patterns.

Middle-out. Forecast at an intermediate level, such as family, then allocate down and sum up. It balances stability and detail, and is a common practical choice.

Modern approaches reconcile forecasts from all levels, using statistical methods that adjust each level to be consistent with the others while minimizing error. You do not need a sophisticated tool to benefit from the idea: forecast at two levels and make them agree.

A worked reconciliation

Suppose the family forecast, built top-down from steady aggregate history and a known price change, is 12,000 units for next month. Bottom-up forecasts for the four SKUs in the family are 5,000, 3,400, 2,000 and 1,000, which sum to 11,400 (illustrative). The two views disagree by 600 units, or 600 / 12,000 = 5 percent. A simple proportional reconciliation scales each SKU by 12,000 / 11,400 = 1.0526:

• SKU 1: 5,000 × 1.0526 = 5,263
• SKU 2: 3,400 × 1.0526 = 3,579
• SKU 3: 2,000 × 1.0526 = 2,105
• SKU 4: 1,000 × 1.0526 = 1,053

Check: 5,263 + 3,579 + 2,105 + 1,053 = 12,000. This keeps the SKU mix from the bottom-up view and the total from the family view. It is the right choice when the family forecast is more reliable, for example because it reflects a confirmed event that applies to the whole family. If instead the bottom-up forecasts reflect specific SKU news (a listing gained on SKU 2, say), scale the family forecast up or choose a reconciliation that gives that SKU more weight. The point is to make the choice deliberately, record the reason and avoid a hidden gap between the two views.

Segment the assortment

Segment SKUs by value and by demand pattern. A common approach is a grid of ABC (by revenue or margin) and XYZ (by variability). A-X items are high value and steady: forecast with seasonal models, review by exception. A-Z items are high value and volatile: give them planner attention and driver-based adjustments. C-X items are low value and steady: automate. C-Z items are low value and erratic: set simple inventory policies, accept the noise and consider whether the item earns its place in the range. This focuses the human work on a small share of the SKUs. See ABC analysis.

Intermittent SKUs

For SKUs with many zero periods, forecast demand size and interval separately, using methods designed for sporadic demand, or forecast at an aggregated level (for example, across locations or weeks) and allocate. For these items, the replenishment policy matters more than the weekly forecast. Review the stocking decision: whether to stock at all, where to stock, and how much. For some slow SKUs, make-to-order or central stocking is better than spreading small amounts across many locations.

New and retiring SKUs

Link new SKUs to analogs and assign them to a family, so that they borrow the family's seasonality and a share of its volume. Plan the transition from the old SKU: reduce its forecast as the new one ramps up, and plan the last buy. For retiring SKUs, plan a run-down and decide when the item leaves the forecast. Allow for the overlap period in which both are sold, since supply chains rarely switch on a single day. Keep a list of items in transition so that none is forecast twice or forgotten.

Location and channel detail

Forecasting at the location level adds another dimension. A common practice is to forecast at the national or regional level and allocate to locations using recent sales share, updated more frequently than the national forecast. Fast-changing locations, such as a new store or a region with a local event, may need a manual override. Where replenishment is centralized, a national forecast with a good allocation rule may be enough. Where there are many forward locations, the allocation rule needs attention, since small errors multiply.

Automate and manage by exception

At this scale, you must automate routine forecasts and review exceptions. Exception rules might flag items whose forecast changed by more than a threshold since last cycle, items whose recent actuals fell outside a prediction interval, items with a stockout or a surge, new items in their first weeks, items with a promotion, and items with persistent bias. Planners work through the flagged list ranked by value. If the list is too long, the thresholds are too tight. If it is too short, you may be missing problems. Tune the rules using history. See planning automation.

Data requirements

SKU-level forecasting needs a clean item master with consistent identifiers and hierarchies, history that links old and new SKUs through transitions, flags for stockouts and promotions, and location attributes. Duplicates, inconsistent units of measure and broken hierarchy mappings are the most common reasons that detailed forecasts fail. A data foundation project is often the first step. See how to build a clean data foundation for planning.

Measure at the right level

Accuracy at the SKU level will always look worse than accuracy at the family level, since noise cancels as you aggregate. Report both. Use a volume-weighted measure and bias, and measure at the lag of the replenishment lead time. Set targets by segment. For intermittent items, use measures designed for them, or evaluate by service and inventory outcomes. See WMAPE.

Common mistakes

1. Forecasting every SKU the same way.
2. Bottom-up everywhere, chasing noise.
3. Top-down everywhere, missing SKU-level shifts.
4. No reconciliation, so the finance and operations views disagree.
5. Ignoring transitions between old and new SKUs.
6. Reviewing everything instead of exceptions.
7. Dirty item master.

FAQs

What is SKU-level forecasting?

Producing a demand forecast for each stock keeping unit, usually by location, to drive replenishment and inventory decisions.

Is top-down or bottom-up better?

Neither universally. Top-down is stable and suits mature families. Bottom-up captures SKU detail but is noisy. Middle-out or reconciled approaches combine the strengths.

How do I forecast thousands of SKUs?

Segment the assortment, automate stable items with suitable models, apply intermittent methods to sporadic items and manage by exception.

How do I forecast SKUs with little history?

Use analogs and family-level patterns, allocate from a higher level and update quickly as data arrives.

How accurate should SKU-level forecasts be?

Less accurate than aggregates. Set targets by segment using your own history and a naive benchmark.