Choice modelling explained: how retailers predict product winners

What is a discrete choice model?

Discrete choice models conduct retail forecasting by modelling the probability of consumer choice. The technique estimates which product a customer will select when presented with multiple options. Choice modelling is a method to model a decision process of an individual in a particular context and can predict how individuals will react in different situations, such as what happens to demand for product A when the price of product B changes.

Predictions of customer choices under different marketing environments can be constructed by choice models, with an example being the predicted purchase probability of a customer for a particular product as a function of its own price or the price of another product. Retailers use these probabilities to optimise assortment, pricing and inventory decisions.

How does conjoint analysis work?

Discrete choice experiments include multiple attributes of a product, such as price, material, production method, brand, certification and social responsibility. Discrete choice modelling asks respondents to consider multiple product profiles at the same time. Researchers then analyse which attributes drive the final choice.

Random parameters logit and latent class models estimate average preferences and segment-specific preference structures. The two most popular models used to predict choice behaviour are the multinomial logit and multinomial probit models. These statistical methods have been standard in retail operations and revenue management for decades, providing necessary input data for inventory and assortment optimisation.

Where does AI improve traditional choice models?

AI-driven forecasting cuts supply chain errors by 20% to 50%, and machine learning demand models reach around 92% accuracy against roughly 75% for traditional methods. The improvement comes from AI's ability to process larger and more complex datasets that traditional statistical models cannot handle efficiently.

Machine learning techniques address several limitations of classical choice modelling. Attribute-based forecasting and lookalike SKU modelling use colour, fabric, silhouette and price tier to identify comparable styles from prior seasons and construct a demand curve based on attribute similarity. This approach solves the cold-start problem for new products that lack historical sales data.

In 2025 Amazon deployed a new foundational AI forecasting model that predicts what customers will want, where and when, for hundreds of millions of products a day, adding time-bound signals like weather and holiday schedules to sales history, and reported the forecasts drove a 10% improvement in long-term national forecasts for deal events and a 20% improvement in regional forecasts for millions of popular items.

What data do AI forecasting models use?

AI models expand beyond the customer surveys and stated preferences that power traditional conjoint studies. Most use cases need at least one to two years of historical sales data so the model can learn seasonality and holiday patterns, and beyond sales history, useful inputs include customer purchase records, inventory movements and promotion calendars.

AI models analyse sales trends, social media activity, economic indicators and weather patterns to predict demand fluctuations. Retailers combine their proprietary data sets with web-crawled publicly available data on Twitter and the Google Search Volume Index. This real-time data integration allows models to detect emerging trends weeks before they appear in sales figures.

How are fashion retailers applying AI choice models?

75% of fashion executives are prioritising AI for demand forecasting, inventory optimisation and cost control. AI-powered forecasting use cases have the potential to reduce inventory by 5 to 15 percent and to achieve a 15 to 25 percent improvement on stockouts. AI forecasting reduces lost sales from stockouts by up to 65%, which represents revenue recovered without additional marketing spend.

In merchandising, AI can automate data consolidation, forecasting and promotional planning to unlock capacity gains of about 40 percent. John Lewis is investing in AI as part of its £800 million transformation programme to push its product catalogue onto AI platforms so customers can find John Lewis products while chatting with AI.

The application extends beyond demand prediction. After prestige beauty group Orveon Global added AI-powered search, merchandising and product recommendations, they reported an immediate 10% to 15% lift in average order value for each brand after the cross-selling functionality went live. Choice modelling principles underpin these recommendation engines by estimating which products a customer is most likely to purchase next.

What are the limitations of AI choice models?

About 71% of merchants say retail AI tools have had limited to no effect on their business so far. Most fail at execution, not analysis. The gap between pilot projects and scaled deployment remains wide across the industry.

Deep learning models are usually treated as a black box; hence, interpretation remains tricky. Merchandisers often cannot explain why a model recommends a particular assortment or price point, which creates trust issues when the model contradicts experience. A technology partner that has built specific models for fashion demand forecasting, trained on fashion-specific data patterns such as seasonality, trend cycles and markdown dynamics, will outperform a generic machine learning platform applied to fashion data.