Skip to content
DT Nova

Predictive analytics for retail: use cases, data and limits

Predictive analytics for retail uses the data a retailer already has, sales, stock, prices, promotions and the calendar, to estimate what will happen next: demand per item and store, the effect of a promotion, the staff a store will need, or the customers likely to stop buying. This guide covers the use cases that pay back, the data each needs, how to tell whether a forecast is any good, and the EU rules that apply when predictions touch individual customers.

The short version

  • Start with demand forecasting for replenishment, where the value shows up directly in stock, availability and waste.
  • Most retail series are sparse. Many items sell zero on most days in most stores, and forecasting methods have to handle that.
  • Recorded sales are not demand. When a product was out of stock, sales understate what customers wanted.
  • Every model has to beat a simple seasonal baseline on your own data, measured in money, not only in accuracy.
  • Predictions about products and stores raise few legal questions; predictions about individual customers raise several.

Retail use cases that pay back

Use caseQuestion it answersDecision it drives
Demand forecastingHow many will sell, per item, store and day or week?Orders, allocation, stock levels
Promotion forecastingHow much extra will a promotion sell, and what does it take from other items?Promotion plans, extra stock
Markdown and clearanceAt which price will remaining stock sell before the season ends?Timing and depth of discounts
Fresh and short shelf lifeHow much will sell before the use-by date?Daily orders, production, waste
Store traffic and labourHow many customers per hour, and how much work?Staff schedules
Customer churn and valueWhich customers are likely to stop buying, and what are they worth?Retention offers, marketing budget
New products and storesWhat will a new item or location sell, based on similar ones?Range, first orders, site choice
Returns and fraudWhich orders are likely to be returned or fraudulent?Checks, sizing advice, policies

Most of these share one foundation: a clean, consistent history of sales, stock and prices at the level where decisions are made. Building that foundation is usually the larger part of the work, and once it exists each further use case costs much less.

Ecommerce predictive analytics

Ecommerce predictive analytics works with richer data than a store: every visit, search, product view, basket and return is recorded, and each order is tied to a customer account. That makes some predictions possible that a physical shop cannot make, and it brings rules of its own. The use cases that come up most in predictive analytics for ecommerce:

  • Recommendations. Which products a shopper is likely to want next, from what similar customers bought and what this one viewed. The baseline to beat is simple: best sellers in the category and "bought together" pairs.
  • Return risk. Which orders are likely to come back, often because of size or fit, so the shop can show sizing advice before the purchase rather than pay for the return after it.
  • Customer value and churn. How much a customer is likely to spend over the next year, and which subscribers or repeat buyers are drifting away, so retention offers go to the people worth keeping.
  • Delivery and stock promises. Whether an item will be in stock and when an order will really arrive, which drives the promise shown on the product page.
  • Demand per product and channel. The same forecasting as in a shop, but with web traffic and marketing spend as inputs, and with marketplaces and the own web shop forecast separately.

Two EU rules matter here. Browsing data is collected through tracking technologies, and Article 5(3) of the ePrivacy Directive 2002/58/EC requires consent to store or read information on a visitor's device unless it is strictly necessary for the service, so a model trained on behaviour only sees the visitors who agreed. And marketplaces that let other sellers offer products are online platforms under the Digital Services Act, Regulation (EU) 2022/2065 (read 11 November 2026): Article 27 requires a platform that uses recommender systems to set out the main parameters in its terms and conditions in plain language, and any options users have to change them, although Article 19 exempts micro and small enterprises from that section. A web shop that sells only its own products is not an online platform in that sense, but explaining why a product is recommended is good practice either way. Prediction about individual customers also falls under the GDPR rules in the section on personal data below.

Demand forecasting in practice

The most studied public example of retail forecasting is the M5 competition, run by the University of Nicosia from 2 March to 30 June 2020 on the Kaggle platform. It used hierarchical sales data made available by Walmart, from single items up to departments, product categories and stores in California, Texas and Wisconsin, with explanatory variables such as prices, promotions, day of the week and special events. Close to 6,000 participants made around 100,000 submissions, and the organisers note that the majority of its 42,840 time series show intermittency: sporadic sales including zeros (University of Nicosia, M5 Competition, read 11 November 2026).

That last point is the practical lesson. At the level of one item in one store on one day, retail demand is mostly small numbers and zeros. Forecasts at that level are noisy whatever the method, so good systems forecast at the level each decision needs, reconcile forecasts up and down the hierarchy, and give a range rather than a single number, because the order quantity depends on how much uncertainty there is, not only on the average. The same store-level forecasts of footfall and transactions can plan staffing too; see retail staff scheduling software.

The data you need

  • Sales transactions at item, store or channel and day level, with returns separated.
  • Stock and availability. Daily stock or at least out-of-stock flags. Without them, a week with no sales because the shelf was empty looks like a week with no demand, and the model learns to order less.
  • Prices and promotions, including the mechanism (price cut, multi-buy, display) and where the promotion ran.
  • Calendar and events: public holidays, school holidays, paydays, local events and, for some categories, weather.
  • Product data: category, attributes and links between old and replacement items, so new products borrow history from similar ones.
  • Online and store channels separately, because they respond to different drivers.
  • Customer data from loyalty programmes or accounts, only for customer-level use cases, and only with a lawful basis.

How to tell whether a forecast is good

Accuracy figures alone mislead. A forecast is judged against a simple baseline, such as the same week last year adjusted for trend, on data it has not seen, over a period long enough to include promotions and holidays. Useful measures are the error weighted by volume, which keeps slow sellers from dominating the average, and bias, which shows whether a forecast is consistently too high or too low. Bias is often the more expensive problem, because it turns directly into overstock or empty shelves.

The final test is the decision. Run the forecast through the ordering rules and compare stock levels, availability, waste and markdowns with what actually happened. A model that is slightly more accurate but does not change orders is not worth running.

Personal data and the EU rules

Forecasts of products and stores use aggregated data and raise few legal issues. Predictions about individual customers are different. Article 22 of the GDPR gives people the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them, and profiling for marketing needs a lawful basis and transparency. Personalised prices have their own rule: Directive (EU) 2019/2161 added a requirement to the Consumer Rights Directive that consumers be informed, where applicable, that a price was personalised on the basis of automated decision-making (both read 11 November 2026). Retailers that use cameras to measure footfall or shelf availability face further rules, covered in our guide to computer vision in retail.

How widely retailers use it

Use of AI is still uneven across companies. Eurostat reports that in 2025, 19.95% of EU enterprises with 10 or more employees used at least one AI technology, against 55.03% of large enterprises (Eurostat, Use of artificial intelligence in enterprises, read 11 November 2026). Statistical forecasting is older than AI and more common than these figures suggest, often inside ERP and replenishment systems, but the gap between large and smaller retailers is real.

Buy a planning suite or build your own?

Retail planning suites from vendors such as RELEX, Blue Yonder and SAS cover forecasting, replenishment and allocation end to end, and for large retailers whose processes match their assumptions they are often the right purchase. A custom model fits better when the problem is specific, such as fresh production, a marketplace with thousands of sellers or an unusual product range, when you want to prove the value on one category before a larger purchase, or when forecasts need to feed systems the suites do not connect to. Either way, the data foundation is the same, and it remains yours.

Common mistakes

  • Training on sales without correcting for out-of-stock days.
  • Forecasting at a finer level than any decision needs.
  • Reporting accuracy without a baseline, or without bias.
  • Ignoring promotions and price changes, then blaming the model for promotion weeks.
  • Building a forecast that nobody uses, because orders are still placed by habit.
  • Starting with customer-level predictions before product-level basics work.

How to start

  1. Pick one decision, such as weekly orders for one category, and one measure of success in money.
  2. Collect two years of sales, stock, prices and promotions for it, and fix the obvious gaps.
  3. Build the baseline first, then one better model, and compare both on held-out weeks.
  4. Run the better one alongside the current process for a few weeks before switching.
  5. Measure stock, availability and waste, then widen to the next category.

How we help

We build the data foundation and the forecasts behind retail decisions, and connect them to the systems that place orders and plan staff. See data analytics consulting for the questions-first approach, inventory management software and distribution resource planning software for where forecasts drive stock, employee scheduling software for labour, and our guide to the types of analytics for where prediction fits among them. For the same methods applied to machines, quality and production, see manufacturing predictive analytics. For customer-level scores in marketing, see predictive analytics for marketers; for prices that follow demand, see AI dynamic pricing.

Questions people ask

What is predictive analytics in retail?

It is the use of historical data, such as sales, stock, prices, promotions and calendar events, to estimate what is likely to happen next: how much of each product will sell in each store, which customers are likely to stop buying, or how many staff a store will need on Saturday. The estimates then drive decisions on orders, prices, staffing and marketing.

What is the most common use of predictive analytics in retail?

Demand forecasting for replenishment. Forecasting how much of each item will sell in each store or channel, and ordering to match, affects stock levels, availability and waste directly, so it is usually where retailers start and where the value is easiest to measure.

How much data do we need?

For weekly or daily demand forecasts, at least one full year of sales history, ideally two or more so that seasonality can be learned, at item and store level, together with stock levels, prices and promotions. Products with very short histories are forecast from similar products until their own history builds up.

Do we need machine learning for retail forecasting?

Not always. A seasonal baseline with promotion and holiday effects is often a strong start, and any more complex model has to beat it on your own data to earn its place. Machine learning models such as gradient boosted trees help when many drivers interact, for example prices, promotions, weather and local events.

Is personalised pricing allowed in the EU?

It is regulated. Since the Omnibus Directive amended the Consumer Rights Directive, traders must inform consumers where a price was personalised on the basis of automated decision-making. Automated decisions with legal or similarly significant effects on individuals are also restricted by Article 22 of the GDPR.

Should we buy a retail planning product or build our own models?

Retail planning suites are a good fit when your processes are close to their assumptions and you have the team to run them. Custom models pay off for specific problems the suites handle poorly, for retailers with unusual products or channels, or as a first step to prove the value before a larger purchase.

Sources