Sales forecasting of motorcycle sales in the Peruvian market

Abstract

This thesis addresses the challenge of accurately forecasting motorcycle sales in the Peruvian market by leveraging advanced machine-learning techniques. It focuses on enhancing sales prediction to improve decision-making and operational efficiency. The study compares three models: Seasonal Autoregressive Integrated Moving Average with Exogenous Variables (SARIMAX), Long Short-Term Memory (LSTM) networks, and Prophet. Utilizing a dataset from January 2017 to November 2023, various data preprocessing techniques, including data harmonization and feature engineering, were employed. Feature selection was conducted using correlation analysis, Random Forest, and Gradient Boosting algorithms. The SARIMAX model achieved a Mean Absolute Percentage Error (MAPE) of 8.06%, LSTM model recorded a MAPE of 9.24% and the Prophet model demonstrated a MAPE of 8.63%. These findings highlight the potential of machine learning models to enhance sales forecasting accuracy. The research contributes to better-informed decision-making, optimized inventory management, and improved operational efficiency in the Peruvian motorcycle market, offering valuable insights for future applications and research.