Choosing Forecast Dimensions: The Battle Between Detail and Accuracy

If you've ever presented a forecast to a sales team, you know the drill. You show a highly accurate forecast at the product family level, and someone immediately asks, "But what about SKU X in the Midwest region?"

This is the "Granularity Paradox" of demand planning.

We all want granular data because it's actionable—you can't ship a "product family" to a "region"; you ship a specific SKU to a specific distribution center. But the more granular you get, the noisier the signal becomes. A forecast that is 95% accurate at the national category level might be 40% accurate at the SKU-store level.

So, how do you choose the right forecast dimensions? Do you forecast by customer to satisfy Sales, or by product to satisfy Operations? And how do you handle the technical reality that 90% of your detailed data points might be zeros?

This guide breaks down the strategy for selecting forecast dimensions that balance statistical stability with business reality.

The Three Core Dimensions of Demand

Before we decide where to forecast, we need to map out the available dimensions. Most demand planning problems exist in a three-dimensional cube: Product, Customer/Geography, and Time.

1. Product Hierarchy

This is usually the cleanest dimension, driven by your master data.

• SKU (Item): The executable level.
• Family/Style: Groupings of similar items (often where seasonality is calculated).
• Brand/Line: Strategic groupings for marketing.
• Category: High-level groupings for budgeting.

2. Customer & Geography Hierarchy

This is where the conflict often starts. Operations thinks in terms of where goods ship (Geography), while Sales thinks in terms of who buys them (Customer).

• Ship-To Location: The physical destination (DC or Store). Critical for deployment.
• Sold-To Account: The paying entity. Critical for sales quotas.
• Channel: E-commerce, Wholesale, Retail.
• Region/Territory: Sales management structures.

3. Time Hierarchy

• Day: Essential for short-term replenishment (especially fresh food).
• Week: The standard for most supply chains.
• Month: The standard for S&OP and finance.
• Quarter: Used for investor guidance and long-term capacity.

Choosing Your "Anchor" Level

You technically can forecast at every intersection of these dimensions (SKU x Store x Day), but you probably shouldn't. You need to pick an "anchor" level—the level where the statistical model runs—and then aggregate up or allocate down.

The Forecastable Level vs. The Execution Level

There is often a gap between where you can predict demand and where you need the number.

• Forecastable Level: The level of aggregation where the demand signal is stable enough to model. Usually higher up (e.g., Product Family x Region x Month).
• Execution Level: The level where you cut POs or schedule production. Usually the lowest level (e.g., SKU x Ship-To x Week).

Middle-Out Forecasting is often the "Goldilocks" solution. You generate the statistical baseline at a middle tier (like Product Sub-Category x Key Account), where the volume is high enough to smooth out noise. You then:

1. Allocate Down: Use historical ratios to break that number down to SKUs for deployment.
2. Aggregate Up: Sum the numbers up to Category/Region for executive S&OP reviews.

This approach gives you the stability of aggregation without losing the detail needed for execution.

The Product vs. Customer Dilemma

One of the most common design questions we hear is: "Should I forecast by Product or by Customer?"

The answer depends on who holds the power in your supply chain.

When to Forecast by Product

If you are a CPG company selling to thousands of small retailers, forecasting by customer is noisy and inefficient. Your demand is driven by consumer trends that happen at the product level.

• Anchor: SKU x Region.
• Use Case: Production planning, inventory deployment.

When to Forecast by Customer

If you are an OEM supplier where 80% of your business comes from 5 key accounts (e.g., Apple, Walmart, Ford), you must forecast by customer. Your demand isn't random; it's negotiated.

• Anchor: SKU x Sold-To Account.
• Use Case: Collaborative planning (CPFR), sales quota management.

The Matrix Approach

Sophisticated organizations often do both. They run a Product Forecast for Operations and a Customer Forecast for Sales. The Demand Planning process then becomes a reconciliation exercise between these two views.

Technical Constraints: The "Sparse Data" Problem

Why not just forecast everything at the lowest level?

Sparsity.

In modern retail and e-commerce, the "long tail" is enormous. If you sell 10,000 SKUs across 500 locations, you have 5 million combinations. On any given day, you might only sell units in 50,000 of those combinations.

That means 99% of your data points are zeros.

Statistical models hate zeros. They create massive variance. A forecast of 0.1 units per day is statistically valid but operationally useless (you can't ship a tenth of a box).

Best Practice: If a combination has intermittent demand (e.g., sells less than once a week), do not forecast it statistically at that level. Aggregate it up to a level where the demand is continuous, forecast there, and use simple logic (like Min/Max settings) for the lower-level execution.

Adaptive Hierarchies: A Modern Approach

Old-school planning systems forced you to pick one rigid hierarchy. You set it up during implementation (Category -> Family -> SKU), and changing it required a database migration.

Modern tools like DemandPlan use Adaptive Hierarchies. Instead of rigid trees, data is tagged with attributes.

• Dynamic Aggregation: You can instantly pivot your view. Want to see the forecast by "Flavor" across all "Regions"? Just select the attributes.
• Virtual Hierarchies: Sales can view the world as Territory -> Customer -> SKU, while Ops views it as Plant -> Line -> SKU. Both views look at the same underlying data, just sliced differently.

This solves the rigid "Product vs. Customer" debate by allowing both views to coexist without duplicating data.

Signs You Have the Wrong Dimensions

How do you know if your current setup is failing? Look for these red flags:

1. Manual Overrides Everywhere: If planners are manually adjusting thousands of SKU-level lines because the statistical model is "glitchy," your forecast level is too low.
2. The "Peanut Butter" Spread: If you forecast at a high level and allocate down by average history, but you keep missing new product launches or regional spikes, your forecast level is too high.
3. Finance/Ops Disconnect: If Finance forecasts in Dollars by Month and Ops forecasts in Units by Week, and they never match, you have a dimensionality mismatch.

Best Practices Checklist

When designing your forecast dimensions, run through this checklist:

• [ ] Align with Lead Times: If your lead time is 12 weeks, your forecast accuracy matters most at the 3-month lag, usually at the monthly bucket level.
• [ ] Don't Forecast What You Can't Influence: If you can't differentiate sales between two similar stores in the same city, aggregate them into a "City" node.
• [ ] Separate the Signal from the Noise: Use Bottom-Up forecasting for short-term execution and Top-Down for long-term strategic planning.
• [ ] Consistency is King: Ensure your history and your forecast use the same dimensions. You can't forecast by "Channel" if your historical sales data doesn't tag the channel.

Conclusion

There is no single "correct" level of granularity. The right choice is a trade-off between the statistical need for data volume and the business need for specific detail.

Start with a "Middle-Out" approach: forecast where the signal is strongest, and use business rules to handle the details. And crucially, invest in tools that allow you to pivot these views dynamically, rather than locking you into a rigid hierarchy from 1995.

---

Ready to stop fighting with rigid spreadsheets? See how DemandPlan's Adaptive Hierarchies let you slice your data any way you need it.