How to Build a Forecast Hierarchy for SKU, Channel, Store, and Category

When we talk about hierarchy forecasting, we get a generic diagram of a single product ladder: SKU rolls up to product, product to category, category to total. Pick top-down or bottom-up, reconcile, done.

That's a useful picture, and it's incomplete in a way that matters. A real consumer brand doesn't forecast along one dimension; it forecasts the same demand sliced four ways at once: by SKU, by channel, by store, and by category. Those aren't four rungs on one ladder. They're four independent dimensions that combine into a grid, and the skill in building a forecast hierarchy is designing that grid well and keeping every view of it consistent, not choosing a single direction to forecast in.

This guide covers what a forecast hierarchy actually is, how to structure it across all four dimensions, which level to forecast at, and how to make every level agree. It's grounded in the established hierarchical-forecasting literature, not vendor folklore, and it treats AI as what finally makes the whole thing practical, rather than a feature bolted on the end.

What is a forecast hierarchy?

A forecast hierarchy is the set of levels at which you produce and view demand forecasts, from the most granular (a single SKU in a single store) up to the most aggregate (total company demand). Retail sales data is naturally hierarchical: bottom-level sales at a SKU-store level add up into categories, channels, regions, and ultimately the total.

The reason a hierarchy exists at all is that different decisions need different levels. A replenishment order needs a forecast at the SKU-store. A production plan needs SKU or product-family. A channel P&L needs channel-level. An S&OP conversation needs category. The same underlying demand has to be visible and trustworthy at every one of those levels, which is exactly where the difficulty starts.

The coherence problem: why levels don't add up

Here is the core problem the entire discipline exists to solve. If you forecast each level independently- a model for SKUs, a separate model for categories, another for channels- the numbers won't reconcile. The sum of your SKU forecasts won't equal your category forecast, which won't tie to your channel forecast, which won't match the total. This is well established: while historical data is coherent by construction (higher levels are literally the sum of lower ones), independently produced forecasts across levels rarely are.

Incoherent forecasts are worse than an academic annoyance. They mean sales is committing to a channel number that the supply plan (built from SKU forecasts) can't support, and finance is budgeting from a category total that doesn't match either. Everyone is working off a different number for the same demand, the exact problem a hierarchy was supposed to prevent. Coherence- every level summing consistently to every other, is the non-negotiable property of a forecast hierarchy, and achieving it is what "reconciliation" means.

SKU, channel, store, category: it's a grouped structure, not one ladder

This is where the four dimensions in the question matter, and where most guides quietly simplify.

A pure hierarchy is strictly nested: each SKU belongs to exactly one category, each category to one total, what the forecasting literature calls the many-to-one rule (many lower nodes group into one parent, and a node never has two parents). Product structure (SKU → product → category) is a clean hierarchy like this.

But channel and store aren't rungs on the product ladder; they're separate dimensions that cut across it. The same SKU sells through DTC, Amazon, and retail; the same category spans every store. When you combine dimensions that each partition the data independently, you don't get a single tree; you get a grouped structure (the literature calls it a grouped time series), effectively a grid: SKU × channel × store, with category as an aggregation of the SKU axis. Total demand can be reached by aggregating along any dimension, and the same bottom-level cell (this SKU, this channel, this store) rolls up in multiple different ways.

Two practical rules follow directly:

• Keep each dimension internally clean (many-to-one). Within the product dimension, an SKU should map to a single category, not multiple. Where your merchandising hierarchy violates this, a product filed under two categories- fix it before you forecast, or the reconciliation math breaks.
• Decide which cross-products you actually plan at. You rarely need every cell of SKU × channel × store; that's a combinatorial explosion of mostly-empty, noisy series. You need the intersections tied to real decisions: SKU-store for replenishment, SKU-channel for channel supply, category-channel for the channel P&L. Design the grid around the decisions, not around completeness.

Getting this structure right is more of the battle than choosing a reconciliation method; a well-designed grid makes the rest tractable, and a bad one makes it impossible.

Which level should you forecast at?

Now the question everyone actually asks, and the reframe that matters.

The classic framing offers three choices. Bottom-up: forecast every SKU-store series and sum upward. It loses no granular information, but the bottom level is noisy and sparse, so aggregate accuracy often suffers, and the compute is heavy. Top-down: forecast the total, then split it using historical proportions. It's stable and cheap, but it loses the individual behaviour of lower levels; a fast-growing SKU gets the average's shape, not its own. Middle-out: forecast at an intermediate level, then aggregate up and disaggregate down.

The mistake is treating this as a single either/or you're locked into. The modern, evidence-backed answer is different: forecast at the level where the signal is cleanest, then reconcile in both directions. Demand at a single SKU-store cell is mostly noise; demand at category or category-channel is mostly signal. Seasonality, trend, and promotional response are far clearer once individual-cell randomness averages out. So you forecast where the signal-to-noise ratio is highest (often an intermediate level such as SKU-region-week or category-store-week), and use reconciliation to push that forecast up to totals and down to the granular cells you need for execution.

The direction of travel isn't a philosophy you pick once; it's chosen per level by where the information actually is. That's why "top-down or bottom-up?" is the wrong opening question. The right one is: at which level is this demand most predictable, and how do I make every other level consistent with that?

How to reconcile: from proportions to optimal reconciliation

Reconciliation is what turns level-specific forecasts into one coherent set. There's a clear progression in sophistication.

Proportional methods are the classic approach: top-down splits a total by each child's historical share; bottom-up simply sums. They're simple and coherent, but crude; top-down forces the parent's shape onto every child, and bottom-up compounds noise upward.

Optimal reconciliation is the modern standard, and it's genuinely better. Instead of trusting one level and deriving the rest, it takes the independently produced ("base") forecasts at every level and adjusts them all simultaneously to be coherent, in the way that minimises overall forecast error. The widely used version is MinT (minimum trace), which uses the forecast-error covariance across the hierarchy to weight the adjustment, so information flows between levels rather than one level dictating to the others. Published retail analyses report that proper hierarchical reconciliation of this kind can improve accuracy by roughly 15–25% over naïve bottom-up or top-down approaches, because a well-reconciled forecast borrows strength across levels: the clean seasonal signal visible at category level gets pushed into the noisy SKU-store forecasts that couldn't see it on their own.

That last point is the one to internalise. Reconciliation isn't just bookkeeping to make numbers tie out; done well, it improves accuracy at every level by letting each level lend its clearest signal to the others.

Where AI changes hierarchy forecasting

Everything above is classical and sound. What AI changes is that it removes the constraint that forced the top-down-vs-bottom-up trade-off in the first place: the assumption that you forecast at one level and derive the rest.

Hierarchical features instead of a single level. A machine-learning model doesn't have to be handed one level's series. It can take features from every level at once: the SKU's own history, its store's pattern, its category's seasonality, its channel's trend, and learn from all of them simultaneously. The granular forecast combines the stable macro signal and the local detail, rather than forcing you to choose which one to forecast and reconcile. This is the same driver-based modelling described in factor contribution in demand forecasting.

New products via the hierarchy. The cold-start problem- no SKU-level history for a launch- is solved by the hierarchy itself: anchor the new SKU on its category trend and comparable products' behaviour, then let its own signal take over as data arrives. That's demand planning for new products expressed as a hierarchy problem.

Coherent probabilistic forecasts. Reconciling point forecasts is well understood; reconciling distributions so that uncertainty is coherent across levels is harder and more valuable, because it lets you set safety stock from a probability distribution at the SKU-store level while keeping the aggregate risk view consistent.

Continuous, automated reconciliation. Instead of a monthly manual roll-up, agentic systems keep every level coherent as new demand arrives automatically, and pick the best model per series - the AutoML-for-planners approach applied across the whole grid rather than one level at a time.

AI doesn't discard the hierarchy; it makes a multi-dimensional one practical to run, where classical methods forced you to simplify to a single ladder to keep it manageable.

A practical sequence for building your forecast hierarchy

1. Map the dimensions to decisions. List the decisions you make: replenishment, production, channel supply, category planning, S&OP, and the level each needs. That list is your required hierarchy; don't build levels no decision consumes.
2. Clean each dimension to many-to-one. Ensure each SKU maps to one category, each store to one region, and so on. Fix a merchandising hierarchy that double-files products before forecasting, not after.
3. Choose the planning intersections. Decide which SKU × channel × store cross-products you actually plan at (SKU-store for replenishment, SKU-channel for channel supply), rather than every combination.
4. Identify the best signal-to-noise level. Find the level where demand is most predictable, often an intermediate category-store or SKU-region level, and make it your primary forecasting level.
5. Reconcile in both directions. Use optimal reconciliation (MinT-style) to push that forecast up to totals and down to execution granularity, so every level is coherent, and each lends signal to the others.
6. Handle new products through the hierarchy. Anchor launches on category and comparables until their own history matures.
7. Automate and monitor. Keep reconciliation continuous and watch coherence and per-level accuracy over time, rather than rebuilding the roll-up each cycle. The onboarding arc is what the first 90 days look like.

This connects directly to how classification and aggregation interact see ABC-XYZ classification) and to the demand-planning complications that a good hierarchy is designed to absorb, including seasonality that only appears at aggregate levels.

Forecast hierarchy FAQs

What is a forecast hierarchy? A forecast hierarchy is the set of aggregation levels at which demand is forecast and viewed, from a single SKU in a single store at the bottom, up through category, channel, store, and region, to total company demand. It exists because different decisions need forecasts at different levels: replenishment at SKU-store, production at product-family, channel P&L at channel, S&OP at category. The defining requirement is coherence; every level must sum consistently to every other.

What's the difference between bottom-up, top-down, and middle-out forecasting? Bottom-up forecasts the most granular level (e.g., SKU-store) and sums upward; it loses no detail but is noisy and heavy at scale. Top-down forecasts the total and splits it down by historical proportions; it's stable but loses each item's individual behaviour. Middle-out forecasts an intermediate level, and both aggregates up and disaggregates down. The modern approach doesn't pick one permanently: it forecasts at whichever level has the best signal-to-noise ratio and reconciles in both directions.

How do you handle SKU, channel, store, and category in one hierarchy? These four aren't a single ladder; channel and store are separate dimensions that cut across the product (SKU→category) hierarchy, so together they form a grouped or cross-product structure rather than one nested tree. Keep each dimension internally clean (each SKU in one category, each store in one region), then plan at the specific intersections your decisions need: SKU-store for replenishment, SKU-channel for channel supply, rather than forecasting every possible combination, most of which are sparse and noisy.

What is forecast reconciliation? Reconciliation is the process of making forecasts across all levels coherent, so the SKU forecasts sum to the category forecast, which ties to the channel forecast and the total. Simple methods use historical proportions (top-down) or summation (bottom-up). Optimal reconciliation, such as MinT (minimum trace), adjusts the forecasts at every level simultaneously using the forecast-error covariance, which both guarantees coherence and typically improves accuracy by letting each level share its clearest signal with the others.

At what level should you forecast demand? At the level where demand is most predictable, the highest signal-to-noise ratio, which is usually an intermediate level (such as category-store-week or SKU-region-week) rather than the noisy bottom or the over-smoothed top. Forecast there, then reconcile up to totals and down to the granular level you need for execution. The old instinct to "always forecast at SKU" or "always forecast the total" is what modern reconciliation is designed to replace.

How does AI improve hierarchical forecasting? AI removes the constraint that you forecast at one level and derive the rest. A machine-learning model can use features from every level at once: SKU history, store pattern, category seasonality, channel trend, so granular forecasts get both macro signal and local detail simultaneously. It also solves new-product cold-start through the hierarchy (anchoring on category and comparables), produces coherent probabilistic forecasts across levels, and keeps reconciliation continuous rather than a manual monthly roll-up.

Why don't my forecasts add up across levels? Because they were produced independently at each level, and independently produced forecasts are rarely coherent, the sum of SKU forecasts won't equal the category forecast produced by a separate model. Historical data is coherent by construction, but forecasts aren't. The fix is reconciliation: rather than forecasting each level in isolation, reconcile them so every level sums consistently, which also improves accuracy when done with an optimal method.

Where TrueGradient fits

TrueGradient forecasts across the full SKU-channel-store-category structure as one system rather than a stack of disconnected models. Machine-learning models use hierarchical, driver-based features from every level at once; forecasts are probabilistic and kept coherent across levels automatically; new products are anchored through the hierarchy, and reconciliation runs continuously rather than as a monthly roll-up, feeding demand planning, inventory optimization, replenishment and allocation, and S&OP from one consistent set of numbers. The channel-specific view is in channel-based demand planning, and the forecasting-plus-inventory pairing is in the dynamic duo.

Typical time to first measurable outcome is 8–12 weeks; see what the first 90 days look like.

Book a demo · Talk to us

Related reading: Channel-based demand planning for omnichannel retail · Factor contribution in demand forecasting · Demand planning for new products in retail · Probabilistic modelling using prediction intervals · ABC-XYZ classification in supply chain management