How to choose the right AI model for business forecasting in retail

One of the areas where retailers can effectively apply AI for is in optimizing user behavior forecasting. This, however, raises key questions: Which model should be chosen, and what criteria should guide that choice? When is it worth deploying complex models, and when is traditional ML sufficient?

Let’s examine the commercial use of AI through a case study delivered by the BaOne team. The project involved personalizing product recommendations in the mobile app of a major grocery retailer with a network of over 900 stores.

The primary objective was to increase profit margins without compromising sales volume. A particular focus was also placed driving the growth of strategic product categories, including private labels (PL) and global exporters (GE).

In the traditional retail purchase funnel, as the customer moves from Stage 1 (New user) to Stage 4 (Explicit context), the company progressively gains more contextual data about the user. At the same time, the complexity and implementation time of AI models decrease.

To build accurate forecasts based on historical data and behavioral factors, it’s important to use models capable of analyzing complex time-series patterns. Common options include:

To predict a customer’s the next purchase, the most frequently applied approaches include sequence modeling, attention mechanisms, and intent modeling.
Examples of models and techniques:

For recommendation systems focused on cart  expansion, approaches often rely on embeddings, multi-task learning, and memory networks.
Examples of models and techniques:

When implementing algorithms, BaOne experts following these guiding principles:

To achieve the retailer’s goals of margin growth, increased turnover, and promotion of PL and GE brands, BaOne experts adopted an approach focused on expanding the customer’s shopping cart by recommending products based on historical context and business rules. This strategy drove quick wins with limited strain on computing resources.

However, as with any AI project, a strong business idea must underpin a successful product recommendation. In this project, BaOne combined the following elements:

Given that maximizing margin ROI was the primary objective, a proven solution was chosen:

Unlike traditional recommendation systems, the focus was not on precisely predicting the next action, but on expanding the user context. This is particularly important when aiming to increase the average ticket size, where key constraints include:

This approach allows for flexible adaptation to shifts in demand and supports the delivery of personalized, commercially effective solutions.

Results

By building on a strong business idea and a hybrid recommendation model, the retailer achieved the following:

The project introduced tools to manage personalized offers via tailored communication channels such as email, mobile apps, push notifications, SMS and more. Additional functionality was developed to support flexible segment management and user activation through RFM-based targeting.

Modern neural network models—such as BERT4Rec, SASRec, or other transformer-based architectures—require significant computing power and complex infrastructure. The decision to adopt them ultimately hinges on striking the right balance between forecast precision, the cost of training and maintenance, and the business impact delivered.

In practice, gains in business performance tend to decrease with each upgrade to more advanced recommendation tools. That’s why the underlying business idea should remain the primary driver in model selection.

As BaOne’s experience shows, hybrid approaches that combine traditional ML-driven recommendations (cost-effective, fast and stable) and complex scenarios powered by neural networks (e.g., personalization for VIP customers or long sessions) often provide the optimal solution. This approach ensures system scalability without overspending and allows flexible adaptation to evolving business needs.

The best approach depends on data volume and complexity, ML infrastructure budget, precision and speed requirements, as well as the business’s readiness for long-term investment. When unsure, start with simple models, track their performance and progressively enhance the system where it yields the highest ROI.