Estimación de ventas mediante un algoritmo de K Means para la predicción basados en Machine Learning.

Abstract

This study investigates the estimation of shrimp feed sales through the combination of the K-Means algorithm and a multiple linear regression model, under a quantitative, applied, descriptive, and predictive approach. The study was conducted using a historical database containing 7,216 records and 38 variables corresponding to commercial transactions carried out between January 2025 and May 2026. Data processing was performed in Python through Google Colab, including inspection, cleaning, handling of missing and outlier values, data type conversion, removal of irrelevant variables, One-Hot encoding, and standardization. K-Means was used for clustering, and the optimal number of clusters was evaluated using the elbow method and the silhouette coefficient, determining that the best configuration consisted of two groups. Subsequently, the Cluster label was included as an explanatory variable within a multiple linear regression model in order to estimate Total USD. The model obtained an R² of 0.9883, a mean squared error of 710,081.00, and a mean absolute error of 485.65. The results showed a high degree of agreement between observed and estimated values, although greater deviations were found in negative transactions or those close to zero. It is concluded that the combination of K-Means and multiple linear regression makes it possible to identify transactional patterns and generate useful estimates to support decisions related to inventories, purchases, dispatches, sales targets, and financial resources, considering the specific methodological conditions of the applied model. The research also compared its results with previous studies on commercial segmentation and prediction, finding similarities in the analytical usefulness of these applied techniques.