Publicação
Multivariate Time Series Forecasting of Sales Volume for the BMW Group: A Machine Learning Approach Outperforming Simple Linear Models and Expert Estimates
| Resumo: | This Master's thesis investigates the potential of machine learning (ML) techniques to outperform simple linear models and expert estimates in the multivariate time series forecasting of sales volume for the BMW Group. The existing literature indicates a gap in knowledge regarding the application of advanced ML techniques for sales forecasting in the automotive industry, specifically considering the BMW Group's unique context. To address this gap, the thesis aims to (1) develop and evaluate an ML-based sales volume forecasting model that captures the intricate relationships among various influencing factors and (2) uncover and incorporate key sales factors through feature engineering. The research employs historical sales data, a Light-Gradient-Boosting-Model (LightGBM) algorithm, and extensive feature engineering to translate crucial sales factors into quantifiable features. The results reveal that the LightGBM model outperforms linear models and expert estimates in terms of accuracy and precision, particularly at the beginning of the month. However, the model exhibits sluggishness towards the end of the month, warranting further investigation. Key findings include the identification of significant features, such as product clusters based on their Product Lifecycle, Google Trend data, and useful Covid-19 features, as well as the uncovering of model weaknesses and data blind spots. This study contributes to the existing body of knowledge in automotive sales forecasting by providing a successful ML-based forecasting model that considers various influencing factors while maintaining explainability. The identified limitations and areas for future research offer opportunities to further improve the model's performance in predicting sales volumes more accurately, ultimately benefiting the BMW Group's sales strategy and decision-making processes. |
|---|---|
| Autores principais: | Mashayekhi, Marian |
| Assunto: | Machine Learning LightGBM Data Pipeline Sales Volume Forecasting Time Series Analysis Process Mining |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso embargado |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | This Master's thesis investigates the potential of machine learning (ML) techniques to outperform simple linear models and expert estimates in the multivariate time series forecasting of sales volume for the BMW Group. The existing literature indicates a gap in knowledge regarding the application of advanced ML techniques for sales forecasting in the automotive industry, specifically considering the BMW Group's unique context. To address this gap, the thesis aims to (1) develop and evaluate an ML-based sales volume forecasting model that captures the intricate relationships among various influencing factors and (2) uncover and incorporate key sales factors through feature engineering. The research employs historical sales data, a Light-Gradient-Boosting-Model (LightGBM) algorithm, and extensive feature engineering to translate crucial sales factors into quantifiable features. The results reveal that the LightGBM model outperforms linear models and expert estimates in terms of accuracy and precision, particularly at the beginning of the month. However, the model exhibits sluggishness towards the end of the month, warranting further investigation. Key findings include the identification of significant features, such as product clusters based on their Product Lifecycle, Google Trend data, and useful Covid-19 features, as well as the uncovering of model weaknesses and data blind spots. This study contributes to the existing body of knowledge in automotive sales forecasting by providing a successful ML-based forecasting model that considers various influencing factors while maintaining explainability. The identified limitations and areas for future research offer opportunities to further improve the model's performance in predicting sales volumes more accurately, ultimately benefiting the BMW Group's sales strategy and decision-making processes. |
|---|