limitedDistribution · Industry Research
SKU-Location Demand Forecasting Must Drive Decisions, Not Just Produce Numbers
A forecast review begins, but the VP of planning still cannot answer a basic question: which SKU-location exceptions require action now? Order history sits in
SKU-Location Demand Forecasting Must Drive Decisions, Not Just Produce Numbers
A forecast review begins, but the VP of planning still cannot answer a basic question: which SKU-location exceptions require action now? Order history sits in the ERP, inventory positions in the WMS, plans in the APS, and commercial changes in spreadsheets or emails. By the time planners reconcile them, service levels, inventory exposure, or margin may already be deteriorating.
Effective SKU-location demand forecasting must shorten that decision cycle. This article explains what fails, why forecast accuracy alone is insufficient, and how to build a governed workflow that routes actionable recommendations.
Why SKU-location demand forecasting remains reactive
The immediate problem is rarely a complete absence of data. It is the effort required to turn fragmented data into a reliable planning input.
A 2024 Wakefield Research survey commissioned by LeanDNA covered 250 supply-chain, planning, and inventory executives. It found that 76% reported lacking a predictive supply-and-demand view. Respondents estimated that workers spent 35% of their time manually entering, tracking, or managing data, while 92% said they made subjective decisions at least sometimes. These vendor-sponsored survey findings are not a universal industry census, but they illustrate a recognizable failure pattern: substantial effort goes into assembling information without consistently producing forward-looking guidance.
At SKU-location level, the complexity compounds. The M5 retail forecasting benchmark contained 42,840 hierarchical series, including 30,490 product-store combinations. Based on Walmart data, it included intermittent daily demand and explanatory variables such as prices, promotions, holidays, and events. That is an illustrative benchmark—not an average shipper portfolio—but it shows why spreadsheet-by-spreadsheet review cannot scale.
Failures typically appear at five connected points:
- Master-data inconsistencies prevent reliable matching across systems.
- Late commercial signals miss the forecasting cutoff.
- Local overrides lack standardized reasons or supporting evidence.
- Hierarchical forecasts disagree across SKU, location, region, and business-unit levels.
- Exceptions arrive without confidence scores, business impact, or a clear approval path.
A forecast becomes operationally valuable only when it identifies the exception, explains the driver, and routes the next decision.
Build a governed path from signals to action
A practical target workflow has four stages:
-
Assemble the planning record. Capture order history, current inventory, open customer orders, prices, promotions, lead times, and relevant event signals from ERP, APS, WMS, customer files, and approved planning communications.
-
Validate and reconcile inputs. Normalize SKU and location identifiers, map fields to the target schema, check completeness and timing, and reconcile conflicts against master data. Low-confidence records should enter an exception queue rather than silently contaminating the forecast.
-
Generate and evaluate forecasts. Run granular forecasts across related series, reconcile the hierarchy, and compare results with an established statistical baseline. The M5 results offer an important lesson: although all top-50 Accuracy competition methods used machine learning, 92.5% of participating teams did not beat the top statistical benchmark on average. Advanced modeling requires disciplined data preparation, tuning, validation, and governance.
-
Translate outputs into controlled decisions. Rank exceptions by service, inventory, cost, and margin implications. Route recommended actions to the appropriate planner, capture approval or override reasons, and deliver the authorized change to the enterprise planning system with traceability.
StarDox Intelligence can provide the automation and decision-intelligence layer around this workflow. Before implementation, planners manually collect files, rekey changes, investigate mismatches, and circulate recommendations through email. After implementation, StarDox Intelligence captures and normalizes relevant inputs, validates them against business rules, coordinates exception handling and human review, and delivers approved intelligence into the APS or ERP. Forecasting remains governed by the company’s models and policies; the surrounding information and decision path become structured and auditable.
Optimize the operating decision, not forecast error alone
Accuracy matters, but it is not the final business objective. A 2025 peer-reviewed study using more than 7,500 M5 demand series found that forecast accuracy and inventory performance do not have a simple one-to-one relationship. The preferred method varied with inventory policy, service targets, lead times, demand intermittency, and the relative costs of holding stock and losing sales. The research used inventory simulations on retail benchmark data, not measured production results.
That distinction changes the design requirement. A planner should not receive only a lower error score. The workflow should explain whether the forecast suggests replenishing inventory, reallocating stock, changing a production plan, or escalating a commercial assumption—and why.
One external implementation shows the potential connection to operating KPIs. In an unnamed large Asian CPG company, McKinsey reported that an analytics-supported planning process followed a prior workflow requiring more than five days for a demand plan and more than two days for a dispatch plan. The reported results included 10–12% greater individual-SKU forecast accuracy, 6–8% lower finished-goods inventory, and 3–5% higher order fill rates. This is one company example without a disclosed control group or full measurement methodology—not an industry benchmark, Stargo result, or guaranteed outcome.
Start with one measurable exception path
Begin with a bounded SKU-location segment and map how a demand signal becomes an approved planning action. Record the current cycle time, forecast accuracy, exception rate, service level, and override path. Then identify where data validation, reconciliation, or approval delays the decision.
That baseline gives the chief supply chain officer or VP of planning a testable question: can one governed workflow improve decision speed and consistency without weakening human control? It is a more useful starting point than deploying another forecast dashboard.
Sources
See ROI in 12 weeks
Stargo users see measurable return and operational profitability gains in just 12 weeks, with non-disruptive implementation in 4 weeks or less.