Methodology for Inventory Management in Neighborhood Stores Using Machine Learning and Integer Linear Programming

Context: Nowadays, inventory management poses a challenge given the constant demands related to temporality, geographic location, price variability, and budget availability, among others. In neighborhood shops, this process is manually done based on experience (the data generated are ignored), which is sometimes not enough to respond to changes. This shows the need to develop new strategies and tools that use data analysis techniques. Method: Our methodology predicts the weekly demand for 14 common products in neighborhood stores, which is later refined based on investment capital. The method is validated using a database built with synthetic information extracted from statistical sampling. For the prediction model, three supervised learning models are used: support vector machines (SVM), AutoRegressive models (Arx), and Gaussian processes (GP). This work proposes a restricted linear model given an inversion and the predicted quantity of products; the aim is to refine the prediction while maximizing the shopkeeper's profit. Finally, the problem is solved by applying an integer linear programming paradigm. Results: Tests regarding the prediction and inventory adjustment stages are conducted, showing that the methodology can predict the temporal dynamics of the data by inferring the statistical moments of the distributions used. It is shown that it is possible to obtain a maximum profit with a lower investment. Conclusions: Our method allows predicting and refining inventory management in a neighborhood store model where quantities are managed to maximize the shopkeeper's profits. This opens the way to explore this approach in a real scenario or to introduce new techniques that can improve its performance.