Optimizing Retail Merchandising Through Transparent Hybrid Sales Forecasting

According to Global Sources, researchers publishing in Frontiers in Big Data have introduced XGB–ANN–Attn, a hybrid forecasting framework that combines XGBoost, an artificial neural network, and a lightweight attention mechanism to improve retail sales predictions while maintaining a level of interpretability. That balance matters on the shop floor: a forecast should be strong enough to guide buying and display decisions, but clear enough for the team responsible for those decisions to interrogate. Before opening hours, the practical question is simple: can a new forecast improve the decisions behind the planogram without becoming another opaque dashboard?

A different route through the forecast

Retail sales data is often organized in large, tabular formats, but it contains complex nonlinear relationships and feature interactions such as pricing dynamics, seasonal effects, and store-specific factors. Conventional statistical methods and standalone machine-learning models may either overlook those relationships or become difficult to interpret as model complexity increases.

The authors argue that many existing methods combine model outputs after the prediction stage, which can dilute critical signals and introduce redundancies. XGB–ANN–Attn takes a different route: it fuses tree-based gradient boosting with neural-network architectures at the feature-embedding level, rather than simply assembling final predictions. Its lightweight attention mechanism, inspired by transformer architecture but optimized for tabular data, is designed to focus dynamically on the most informative features.

For merchandising teams, that distinction is more important than a technical name. We are not being handed a mysterious number to obey; we are being given a way to examine which sales signals the system is treating as important before deciding what earns endcap space, how a planogram should change, or where a display deserves more attention. The framework is also presented as suitable for deployment in large-scale retail analytics environments, a key requirement when a useful pilot has to work across more than one store.

What the reported test covered

The researchers tested the model on BigMart and Walmart sales data, two established retail datasets with diverse product assortments, store types, and sales patterns. The comparison included traditional linear regression, standalone XGBoost, CatBoost, LightGBM, TabNet, and plain artificial neural network architectures.

On the BigMart dataset, XGB–ANN–Attn recorded an RMSE of 0.1584 and an R² of 0.9946. On the larger and more complex Walmart dataset, it recorded an RMSE of 0.8652 and an R² of 0.9568, outperforming the comparative models in the reported test.

Those results are strong enough to justify attention, but they are benchmark results from the datasets used in the study—not a promise that the same figures will appear in your own stores. A retailer’s useful KPI is not simply the most impressive R². The more important check is whether the forecast improves the decisions your team makes when planning a display, adjusting an assortment, or responding to a sales trend.

Make the model earn its shelf space

Start with one clearly defined planning question. For example, can pricing dynamics, seasonal effects, or store-specific conditions explain enough of the sales pattern to change what you place on an endcap or how you build a planogram? That gives the forecast a job to do and prevents a sophisticated model from becoming an expensive reporting exercise.

Run the current process and XGB–ANN–Attn against the same retail sales data, then record the RMSE and R² for a direct comparison. Ask both the buyer and the floor manager to review which features the attention mechanism is treating as informative. If the people responsible for the decision cannot follow the logic, the forecast is not yet ready to guide customer-facing space, no matter how polished the dashboard looks.

The go/no-go test should be short: the model should lower forecast error against the current baseline, deliver a strong R² on your own data, and make the decision path clearer. If it cannot improve both the forecast and the practical response, it has not earned shelf space in the planning process.