James Theuerkauf is an industry leader in machine-learning based demand forecasting & inventory optimization. He co-developed the LDSM.
Most retail demand is fairly predictable from far enough away. National weekly denim sales move with the calendar, promotions and the weather, and a competent statistical model will track them. But inventory decisions are made for one product, in one size, at one store, in one week.
At that level, history thins out fast. A planner deciding how many size eight boots to send to one store is often looking at mostly zeros. The M5 “Accuracy” competition, which used Walmart unit-sales data, confirmed that most item-store histories are intermittent: long runs of zero demand broken by the occasional sale.
After several years building forecasting systems for large apparel and footwear retailers, I treat this as a representation problem more than a time-series problem: describing products and stores well enough for a model to learn from all of them at once. The approach that has worked best for me is a multi-tower neural network built on learned embeddings.
Why Item-By-Item Forecasting Breaks Down
Take, for example, a running shoe that has sold in black for three seasons and is launching in green. The green version has no sales, but its style, its subclass and every store have years of history.
A traditional setup forecasts each item in each store from its own sales alone, discarding all of that, and leans on a planner’s hand-picked “like item” for new products.
Global models such as Amazon’s DeepAR moved the field forward by training one model across thousands of items. The open question is how it should describe products and stores.
What Products And Stores Have In Common
Retailers record most of what they know as labels, such as style, store, subclass, color and size, often with thousands of values each. Data scientists call these high-cardinality categorical variables. The standard technique, one-hot encoding, treats every value as unrelated: Style 40213 says nothing about style 40214, even if they are the same jacket in two colors.
Entity embeddings, described by Cheng Guo and Felix Berkhahn, work differently. The model learns a short numerical profile for each style and store, like coordinates on a map. Products that sell alike end up close together, as do stores that behave alike.
In my experience, store profiles routinely pick up patterns nobody entered, such as climate, tourist versus local traffic or a lean toward smaller or larger sizes. Describing both sides this way lets the model forecast a product in a store where it has never sold.
Separating What, Where And When
A multi-tower design borrows from recommendation systems, where one part of the model, or “tower,” learns about customers and another about items before matching them. Here, each tower studies one kind of information before the model combines them. I recommend at least three:
• Product Tower (What): Style and item profiles, hierarchy, fabric, fit, price tier and life cycle stage
• Location Tower (Where): Store or warehouse profiles, market, climate, channel and format
• Context Tower (When): Calendar, holidays, promotions, mark downs, stock on hand and recent sales
Separate towers force the model to explain demand through product, place and circumstance instead of memorizing each item-store pattern. That makes “cold start,” forecasting items with no history, manageable. The green shoe has never sold, but the model has seen the black version, similar shoes and every store.
Letting New Products Borrow From Their Families
Items roll up into styles, subclasses and departments; stores roll up into markets and regions. The usual approaches split a high-level forecast down by historical percentages, or forecast every level and force the numbers to add up with methods such as MinT reconciliation. Both treat the hierarchy as an accounting rule.
Learned profiles let it act as a starting assumption instead. A new cardigan is assumed to sell like its subclass until its own sales say otherwise, while an established item keeps its own pattern. Technically, sparse items are regularized toward their parents: The model is penalized for letting a thinly sold item drift from its family without evidence.
Sizing is where I have found this matters most. Many retailers use one fixed size split per department, yet the right mix depends on fit, category and a store’s customers. A learned model can propose a size split for a new item from similar items in similar stores.
Measuring Accuracy Where It Counts
Errors cancel when you add them up: Too many mediums in one store and too few in another can net to a perfect total. A model can post strong category-week accuracy, often measured as WMAPE (weighted mean absolute percentage error), and still send the wrong sizes to the wrong stores.
I measure accuracy where inventory is committed, split by level of detail, item age and sales volume, and I track bias: whether the model leans consistently high or low.
Mistakes don’t cost the same both ways. Under-forecasting a launch empties shelves when customer response is most telling, while over-forecasting XXS or XXL ends in markdowns. Forecasting a range, such as a low, likely and high estimate (statisticians call these quantiles), and scoring it with pinball loss, which penalizes over- and under-forecasting differently, makes those trade-offs explicit. It also lets ordering systems hold more safety stock where a stockout costs most.
What This Means For Your Organization
Effort shifts from maintaining thousands of individual forecasts and “like item” rules to deciding how the model understands the business. Three questions for leaders:
• Do new products draw on everything we know, or on one comparable a planner picked?
• Do we measure accuracy, including bias, at the level where we buy and allocate?
• Do our forecasts express uncertainty, and does replenishment use it?
Retail planning has long forced a choice between forecasts reliable enough to trust but too broad to act on, and forecasts detailed enough to act on but too thin to trust. A model that learns how products, stores and circumstances relate narrows that gap. Planners still make the calls.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?