Optimizing Pricing Strategies Using Machine Learning: A Prototype Development for Derome

Abstract

This project explored the application of machine learning models to analyze and forecast sales performance using internal transaction data from Derome, a Swedish company. The study implemented clustering, classification, and regression techniques to build interpretable, high-performing prototypes aimed at enhancing decision-making and revenue forecasting.

For clustering, k-means was applied to annual aggregates of product groups based on sales volume and outcome. Using the elbow method, three distinct performance tiers were identified, validated by a high silhouette score of 0.89. These tiers were assigned to transactions and used as labels in a classification task. Among the evaluated models, a decision tree classifier achieved the highest accuracy of 85.86% on unseen data for the year 2024 and outperformed a strong baseline classifier. Feature importance analysis revealed that the supplier was the most influential factor in predicting a transaction’s associated performance tier.

Regression modeling focused on predicting monthly sales using two approaches: forecasting based on individual transaction-level predictions, and direct forecasting from aggregated historical data. Extreme gradient boosting regressor demonstrated the best overall performance, achieving a mean absolute percentage error (MAPE) of 8.96% between the predicted and true values for 2024 using the second approach. It was also used to forecast 2025 sales, achieving a MAPE of 8.07% when compared to the company’s predicted revenue for that year, indicating strong alignment with business expectations.

Despite limitations such as a restricted hyperparameter search, exclusive reliance on internal data, and initial gaps in domain knowledge, the project successfully delivered functional prototypes. These systems provide a foundation for future development and demonstrate the practical value of machine learning in improving business forecasting and decision-making.

Overall, the project met its objectives by uncovering customer behavior patterns, optimizing internal workflows through machine learning, and supporting pricing and strategic planning with interpretable, data-driven models.