Our client, a specialized seafood company was facing tremendous pressure for inventory and network optimization to reduce operational costs and free up working capital tied up in inventory investment. Under the mixed “Make & Buy” strategy, the company operates a supply network optimization consisting of 7 plants and 6 distribution centers in North America (“Make”) and imports a large number of products in full containers from its suppliers in Southeast Asia (“Buy”). With declining market demand, increasing supply complexity and constraints, and a “one size fits all” inventory optimization strategy, they had suffered from sub-optimal inventory performance compared to relevant consumer goods industry benchmarks. Clarkston was engaged to develop an optimal approach for right-sizing inventory targets that would enable this client to achieve tangible inventory cost savings while maintaining or improving service levels.
The Clarkston team started the analysis based on the traditional safety stock calculation method set forth in “Crack the Code”, a well-known article regarding statistical inventory modeling published by MIT. This “MIT Model” has been widely adopted across industries, and many inventory and network optimization models and software packages employ the same or similar formulas. However, the Clarkston team quickly recognized that one significant limitation of the MIT Model could artificially inflate the safety stock requirements: the MIT Model does not take forecast accuracy into consideration and treats all demand variations – both predictable and random – as random. This limitation was particularly problematic for the client, as many of its products are heavily promotion-driven and seasonal, and much of the resulting demand fluctuations are anticipated and reflected in demand forecasts.
The concept of incorporating demand predictability into statistical safety stock calculations itself is not entirely new. Some supply chain professionals have attempted to use forecast accuracy measures (e.g., 1 – MAPE) while others have tried the standard deviation of forecast errors. However, before we started the project, there had not been a well-defined and -reasoned “new model” in the market that makes statistical sense. In addition, it was not clear in which cases the traditional model or the proposed new model should be used. Traditionally, a somewhat arbitrary threshold was often selected to segment products into “good”, “fair” and “poor” forecast-ability groups (e.g., forecast accuracy ≥90% is “good”). But in real-world applications, convincing clients that 89% forecast accuracy justifies a materially different model compared to 90% often proves to run counter to the client’s intuition.
The primary objectives for this project were to develop an industry-first approach and logically incorporate forecast-ability into the statistical safety stock calculations. The goal was to quickly cleanse and analyze large volumes of historical data down to the SKU and location level, while also rationalizing the model selection decision process. Some of the resolutions included developing a proprietary application and a safety stock model selection process that proved to be more objective and intuitive for the company.
Ultimately, the client project was a success on more than one level. Beyond the immediate value delivered to them, perhaps more importantly and strategically, this was a pioneer project in which Clarkston implemented what we believe to be an industry-first statistical analysis that represents a significant improvement upon the traditional MIT Model.
Co-author: Fang Liu