Demand forecasting plays a critical role in organizational planning, encompassing inventory management, capacity allocation, and financial decision-making. However, achieving accurate forecasts can be challenging, particularly in industries characterized by high demand volatility, such as semiconductor assembly equipment manufacturing, exemplified by Besi. Leveraging machine learning (ML) techniques presents a promising solution for effectively forecasting the seasonal and cyclical demand fluctuations experienced by Besi.

Besi, full name BE Semiconductor Industries N.V., is an multinational semiconductor equipment manufacturer which originates from The Netherlands. The company was founded in 1995 by Richard Blickman and now has operations in, among other countries, China, Switzerland and Malaysia. Besi develops leading edge assembly processes and equipment for leadframe, substrate and wafer level packaging applications in a wide range of end-user markets including electronics, mobile internet, cloud server, computing, automotive, industrial, LED and solar energy.
%more about besi

This paper investigates the challenges organizations face when implementing ML models into their demand forecasting processes, aiming to design a framework and implementation plan to guide organizations in adopting ML techniques. The research methodology employed is design science research (DSR), which focuses on developing and validating new designs within existing systems. The paper follows the iterative steps of DSR, including problem identification and motivation, objective definition, design and development, demonstration, evaluation, and communication. This iterative process facilitates collaboration with literature and industry experts to design a practical solution.

The study draws on literature research and exploratory discussions with Besi employees, emphasizing five key areas for investigation: the current forecasting process, existing forecasting techniques, organizational requirements, limitations, and input-output considerations. The findings highlight that Besi, similar to other organizations, employs a multi-layered forecasting process, with the most effective layer for implementing improvements and ML models being the initial forecast. Additionally, Besi predominantly relies on judgment-based forecasting techniques, making the implementation of a neutral ML tool necessary to create a hybrid forecasting system that mitigates human bias. Besi possesses the necessary prerequisites for effective ML techniques, such as clean and abundant data, but lacks the expertise required to construct and implement accurate models. Furthermore, Besi desires a neutral model that counters human bias and inputs historical monthly sales data, with the output expressed as total monthly sales.

To facilitate successful implementation within Besi's forecasting chain, several aspects are explored: the process framework (including integration, monitoring, updating, forecasting, and communication), peripheral considerations (e.g., legal and end-user trust), and the dashboard. The process framework is designed based on the existing forecasting process at Besi, incorporating the ML model, a dashboard, and revised information flows. The steps align with literature recommendations for ML-based forecasting, indicating that initial implementation is expected to face minimal resistance, monitoring is an ongoing task for the forecaster, updating involves improving the model and frequent training with new data, forecasting remains with the same personnel but incorporates additional information sources, and communication remains unchanged.

Peripheral matters, such as regulations and end-user trust, are limited in their impact, with research indicating that no laws impede ML model implementation, while a dashboard can enhance trust among direct users. Gradual implementation in phases, where the model does not hold authoritative power, facilitates organizational acceptance.

Based on insights from all sources, a step-wise plan for ML model implementation is proposed. The initial phase involves assembling the necessary infrastructure, data, and stakeholders, followed by creating and implementing a minimum viable product as a confirmation tool alongside the existing forecasting process. The minimal viable product is a functional model that provides usable accuracy and lists expected monthly sales based solely on historical data and observed fluctuations.

The second and final step focuses on refining the model and presenting its findings through a dashboard, incorporating information from other relevant sources to support the forecaster in making informed forecasts. This phase also enables improved communication with other relevant departments. Ultimately, the forecast incorporates real-time sales data for increased accuracy.

Feedback from Besi representatives indicates that this implementation approach is suitable for their business. The framework and steps presented in this study have been generalized to benefit other organizations, making an academic contribution in the field of ML-based demand forecasting. This contribution stems from the building upon the theory by Caniato in implementing a quantitative forecasting method. The new findings show that there are multiple ways to implement a quantitative method and that, according to this paper, the implementation is best cut up into two phases for smooth transition and maximum acceptance.