MACHINE LEARNING IN RETAIL AND MARKETING

Introduction. This research focuses on addressing the scientific and practical task of implementing machine learning (ML) algorithms into marketing and pricing processes. The relevance of the study is driven by the inability of traditional approaches to effectively process growing volumes of data and facilitate a transition to proactive and predictive business strategies. The aim of the work is to investigate the practical application of ML algorithms for solving two key tasks: pricing optimization through sales forecasting and enhancing marketing strategy efficiency by evaluating the effectiveness of promotional campaigns. The paper presents a methodology for building and training models, along with an analysis of their applicability based on retail trade data.

Materials and Methods. The study focused on historical data from a retail chain, including daily sales, customer traffic, and marketing activities. The study covered data from 1,115 stores, with a total sample size of over 1 million observations. The study employed a combination of machine learning techniques, including regression analysis for sales forecasting and binary classification for evaluating the effectiveness of marketing campaigns. The research architecture is based on the sequential application of gradient boosting algorithms - CatBoostRegressor for the regression task and CatBoostClassifier for the classification task. The pandas library was used for data processing, and feature engineering involved generating time-based features and calculating the performance of stocks based on comparisons with baseline sales. The models were validated using cross-validation and a test set divided by time intervals.

Results. As a result of the study, two machine learning prediction models were developed and tested. The sales prediction model achieved a determination coefficient of R² = 0.837 with a root mean squared error of RMSE = 1254.38, indicating a high accuracy in predicting daily turnover. The stock performance classification model demonstrated balanced accuracy with an F1-score of 0.65 and revealed a significant difference in the effectiveness of promotions between different store types, ranging from 73.6% to 99.7%. It was found that effective promotions lead to an increase in average sales by 87% compared to days without promotions (8244.31 vs. 4406.05). Feature importance analysis identified key influencing factors: number of customers (42.8%), distance to a competitor (15.4%), and store type (12.5%) for the effectiveness of promotions; store ID, day of the week, and the fact of a promotion for predicting sales.

Discussion and Conclusion. The practical significance of the study lies in the creation of a toolkit for optimizing marketing strategies and inventory management in retail chains. The implementation of the developed models allows for a transition from reactive to proactive sales management, improving the accuracy of demand forecasting by 15-20%, and optimizing the allocation of marketing budgets by focusing on the most effective promotion channels. The results obtained demonstrate a significant dependence of the effectiveness of promotions on the type of store and its competitive environment, indicating the need for a differentiated approach to marketing activity planning. Prospects for further research include the development of dynamic pricing systems, the integration of external factors (seasonality, macroeconomic indicators), and the creation of recommendation systems for selecting optimal stock parameters. The developed methodology can be adapted for other retail segments and service industries.

Ключевые слова: машинное обучение, ценообразование, прогнозирование продаж, маркетинг, тестовая выборка, торговля.

Keywords: machine learning, ricing, sales forecasting, marketing, test sample, trading.