Forecasting inventory demand for a semiconductor manufacturer : a case study using machine learning and other methods applied to time series data
Piedrafita Acin, Victor Manuel (2023)
2023
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2023121737970
https://urn.fi/URN:NBN:fi:amk-2023121737970
Tiivistelmä
We have witnessed years where semiconductor scarcity has been repeatedly reported in mass media as a major threat for several industries supplier chains causing limitations in production of goods and indeed prices to rise for final customers. In this challenging environment, proper inventory management can save thousands of euros and can be the difference between failed and successful commercial operation. A key part of inventory management is the demand forecasting. Forecasting methods have traditionally been utilized in this task by professionals, but the arrival of machine learning and other advanced time series models might bring a new approach to provide more reliable and accurate tools in demand forecasting in inventory.
Current research aims to analyze different time series models applied to inventory of a semiconductor manufacturing company and find correlations between part demand patterns and best-performing forecasting models. The literature review focused on the topics of inventory management and time-series forecasting families: classical based on statistics, machine learning, deep learning and other algorithms. As preparatory work for the empirical part, research framework was defined by selecting mature models in each family and researching suitable selection criteria to evaluate the best performing model. The empirical research was a comparative case study of inventory parts and time series models, evaluated with accuracy indicators.
Conclusions from the empirical research confirmed the existence of a triple bottom line in forecasting defined by data nature, models feature, and forecast needs where only accurate results can be reached with balanced elements. For this study case, in spite of stringent forecast requirement and poor “forecastability” in data, gradient boosting showed its adaptability and potential to continue working in the development of inventory tools for this semiconductor manufacturer.
Current research aims to analyze different time series models applied to inventory of a semiconductor manufacturing company and find correlations between part demand patterns and best-performing forecasting models. The literature review focused on the topics of inventory management and time-series forecasting families: classical based on statistics, machine learning, deep learning and other algorithms. As preparatory work for the empirical part, research framework was defined by selecting mature models in each family and researching suitable selection criteria to evaluate the best performing model. The empirical research was a comparative case study of inventory parts and time series models, evaluated with accuracy indicators.
Conclusions from the empirical research confirmed the existence of a triple bottom line in forecasting defined by data nature, models feature, and forecast needs where only accurate results can be reached with balanced elements. For this study case, in spite of stringent forecast requirement and poor “forecastability” in data, gradient boosting showed its adaptability and potential to continue working in the development of inventory tools for this semiconductor manufacturer.