Red parcel lockers with a digital screen and keypad

Omnichannel Demand Forecasting: Channels, Levels and Returns

Omnichannel demand forecasting predicts what customers will buy in each channel (stores, the online shop, click and collect and marketplaces) and where that demand will be fulfilled. A single total forecast is not enough, because the same units can be sold online and picked in a store, shipped from a DC or returned to a different location. A good omnichannel forecast models each channel, the way the channels take sales from each other and the returns that flow back, at the level where replenishment and allocation decisions are made.

Why a Total Demand Forecast Falls Short

A retailer can forecast its total sales of an item well and still put the stock in the wrong places. Channels differ in two ways.

First, the demand patterns differ. Online sales of an item follow different weekly and daily curves from store sales, react to site traffic, search ranking and delivery promises, and cover a wider assortment, including sizes and variants a store does not stock. Store sales depend on local footfall, shelf presentation, weather and the store’s own range. A model fitted to the combined history blends two patterns and describes neither.

Second, demand and fulfillment happen in different places. An online order placed by a customer in one city may be picked in a nearby store, shipped from a regional DC or collected at a third location. For planning, the question is where each unit of demand will consume stock. The channel that takes the order and the node that fills it are separate attributes of the forecast, and both matter.

So an omnichannel forecast needs two splits: demand by channel, so each pattern is modeled on its own drivers, and demand by fulfillment node, so replenishment and allocation see the load each location will carry.

Store and Online Cannibalization

Channels share customers, so they take sales from each other. Opening online sales in a region, adding click and collect, running an online-only promotion or closing a store moves demand between channels as well as creating new demand. A forecast that treats each channel as independent will count the same customer twice after such a change.

A few practical rules help:

  • Forecast channel shift explicitly when the cause is known, such as a new delivery area or an online-only campaign, and reduce the affected store forecasts by the volume the model expects to move.
  • Treat promotions as cross-channel events. A deep online discount can lower store sales of the same item, and a store promotion can pull demand from the web.
  • Watch the total as well as the parts. If channel forecasts rise together after a change, check whether total demand really grew or one channel simply took share from another.

Research supports forecasting the channels together. A study of about 24,000 products at a French cosmetics retailer found that forecasting the online and store channels jointly was beneficial, and that basket data, which shows which products are bought together or instead of each other, improved forecast accuracy and stock control (Omar et al., International Journal of Production Economics, 2023).

Click and Collect, Ship-From-Store and Returns

Hybrid fulfillment changes which stock a sale draws on. Click and collect and ship-from-store orders are online demand that empties store shelves. If store replenishment only forecasts walk-in sales, the store runs short of the items online customers reserve there. Those orders belong in the forecast of the store that fills them, as a separate stream with its own pattern, so replenishment covers both.

A retailer can also hold a protected quantity in each store for walk-in customers, so online orders do not take the last units off the shelf. The size of that buffer should come from the forecast error of each stream rather than a flat rule.

Returns run the other way. They add stock back, often at a different location from the sale, and with a delay. In categories such as apparel and footwear, online return volumes can be high, and buy-online-return-in-store sends those units to a store rather than a DC. Two forecasts then matter:

  • Gross demand: what customers order, which drives picking, delivery capacity and service measures.
  • Net demand: what they keep, which drives inventory investment and the buy.

The return forecast itself depends on product category, season, price point and events such as sales periods. Returned units also need a decision: back to stock in the store, back to the DC, or out to an outlet channel, depending on the business rules for resale.

Which Level Should Omnichannel Demand Forecasting Use?

Retail demand sits in a hierarchy: product (style, item, size), location (store, cluster, region, fulfillment node), channel and time (day, week, season). There is no single right level for the whole hierarchy. The practical rule is to forecast where the signal is strong enough to model, then reconcile to the level each decision needs.

DecisionLevel it needs
Buy and production ordersItem or style by week, total across channels, net of returns
Initial allocation of new-season stockItem and size by store cluster or store, by week
Store replenishmentSKU by store by day, including click and collect and ship-from-store
Online fulfillment center replenishmentSKU by fulfillment node by day
Delivery and picking capacityOrder lines by node and channel by day or hour

There are three common ways to reconcile the levels. Bottom-up forecasting models each SKU, location and channel and adds them up, which keeps local detail but is noisy for slow items. Top-down forecasting models a total and splits it by historical shares, which is stable but misses local change. Middle-out forecasting models a middle level, such as category by store or item by channel, and goes both up and down from there. The methods can be mixed: fast items are forecast bottom-up, slow and intermittent items at a higher level and then split.

Measure accuracy at the level of each decision. A forecast can look accurate in total while store-level forecasts are biased in opposite directions by channel. Tracking error and bias by channel and by fulfillment node shows where the plan will fail.

How the Forecast Feeds Allocation and Fulfillment

The forecast is an input to decisions that move stock. Allocation uses the forecast by store and channel to decide how much of a delivery each location receives and how much stays central. Our post on omnichannel inventory allocation covers shared pools, protected quantities and central versus deployed stock. Replenishment uses the forecast by fulfillment node and its error to size orders and buffers. Order routing then decides, order by order or in batches, which node fills each online order; our post on dynamic and distributed fulfillment compares the two approaches.

Each step changes the next forecast. Routing more online orders through stores raises the store fulfillment stream, and a markdown in one channel shifts demand in the others. Short-term signals such as POS, web traffic and weather help correct the forecast between planning runs, as described in our post on demand sensing.

How Solvoyo Forecasts Demand Across Channels

Solvoyo is a supply chain decision automation platform. Its demand planning forecasts by SKU, channel and location, and selects the best-fit model for each combination from a range of statistical and machine learning methods. The models take in promotions, special days, weather, competitor factors and, for online demand, site visits and click rates. Similarity-based learning covers new items and store openings. In omnichannel planning, the same channel-level forecast feeds store replenishment and allocation. Virtual safety stocks protect store inventory for walk-in customers, and expected returns, predicted from product category, time of year and special events, adjust the inventory recommendations.

Because each store order already counts the online orders that store will fill and the returns it will receive, it matches the stock the store will actually draw on: over 95% of Solvoyo’s plans execute untouched. To see a channel-level forecast on your own data, request a demo.

Demand forecast diagnostics in the Solvoyo platform

Frequently Asked Questions

What Is Omnichannel Demand Forecasting?

Omnichannel demand forecasting predicts customer demand for each sales channel, such as stores, the online shop and click and collect, and for each location that fulfills it. It accounts for how channels take sales from each other and for returns, so replenishment and allocation plan the right stock at each store and fulfillment center.

Should You Forecast Online and Store Demand Separately?

Model them as separate streams, because their patterns and drivers differ, but forecast them together so a sale that moves from one channel to another is not counted twice. Then reconcile the streams to the level each decision needs, such as store and day for replenishment or total by week for the buy.

How Do Returns Affect Omnichannel Demand Forecasts?

Returns add stock back with a delay and often at a different location from the sale. Planners forecast gross demand for picking and delivery capacity, and net demand after returns for inventory and buying. The return forecast depends on product category, season, price point and sales events.


See It Run on Your Own Data

See how Solvoyo turns your supply chain plans into decisions that run on their own.

Solvoyo mascot