New Data-Driven Methods for Managing Complex Inventory Systems in Global Supply Chains

生产各种商品（包括消费品（如保健、个人和美容护理、食品和饮料）、药品、化学品和电子产品）的全球公司正面临三重威胁：(1)为越来越不耐烦和挑剔的消费者群体提供越来越多的产品品种；(2)由于创新速度加快而缩短的生命周期；(3)由于新的在线竞争者的激烈竞争而缩小的利润率。对数百个地点的数千种商品进行计划（并适当储存）的能力已成为业务中越来越重要的方面，既可以获得收入（通过不因库存短缺而错过销售），也可以避免未售出材料的成本（包括环境成本）。与过去有足够数量的数据可用，规划范围更稳定的情况不同，今天这些决策需要在更频繁的基础上使用较少数量的历史数据（尽管在大量产品、地点和时间之间存在复杂的相互作用）。这个研究项目将产生新的方法来帮助这些公司，把以前无法获得的方法从应用数学、统计学和工程学的其他领域发展出来，并成功地应用于这些领域。研究人员与几家公司密切合作，并有代表性的数据，他们可以测试新方法的性能。在当今竞争激烈的全球供应链的复杂数据环境中，有效的库存管理（通过一系列令人眼花缭乱的产品、卓越的服务水平和低廉的价格努力取悦不耐烦的客户）对于各行各业的商业成功至关重要。在产品可用性上的竞争已经成为一个关键的战略武器。这需要新的数据驱动的方法，而不是依赖于分布假设的方法（这是当前实践中的主流）。有五个方面使数据变得复杂。首先，有限的购买历史使得我们很难对已经不确定的消费者有一个完整的了解；因此，除了数据的不确定性外，还必须考虑模型中的参数不确定性。其次，商品是成组购买的，或者作为替代品，因此需要一个可处理的多变量模型，该模型可以用有限的数据构建。第三，由于消费趋势，购买行为存在时间方面的自相关性和滞后性。第四，不是每件物品都在每个时期被购买，也不是以固定的金额购买；除了购买之间的时间和购买数量之间的相关性之外，这种零星的（间歇的）需求建模也很重要。最后，由于季节性和生命周期阶段，购买行为是非平稳的。研究人员将使用一个考虑参数不确定性的框架，一个既稀疏又通用的多变量表示，并通过在数据生成过程中做出最少的假设，从可用的有限数据开始开发快速算法。先前的工作在处理单一产品，有限的历史情况下非常成功，没有间歇性需求，早期测试显示间歇性数据的性能非常好。除了研究之外，研究人员还计划为教育工作者提供工具，以便这些工具可以成为高年级本科生和硕士（包括技术mba）的下一代教学工具的一部分。

Global companies that produce a variety of goods (ranging from consumer products (such as items in health, personal and beauty care, food and beverages), pharmaceuticals, chemicals and electronics) are facing a triple threat: (1) increasing variety of products for an increasingly impatient and discriminating consumer base, (2) shorter life cycles due to increasing speed of innovation and (3) shrinking profit margins due to intense competition from new on-line competitors. The ability to plan for (and stock appropriately) the thousands of items across hundreds of locations has become an increasing critical aspect of the business, both to capture revenues (by not missing sales due to shortage of inventory) as well as to avoid costs of unsold material (including the costs on the environment). Unlike in the past where sufficient amount of data was available, and the planning horizons were more stable, today these decisions need to be made with low number of historical data (albeit with complex interactions across large number of products, locations and time) on a more frequent basis. This research project will yield new methods to help these companies by bringing to bear previously unavailable methods developed from, and successfully applied in other areas of applied mathematics, statistics and engineering. The investigators have worked closely with several companies and have representative data on which they can test the performance of the new methods.Effective management of inventories in today's complex data environment of highly competitive global supply chains (fighting to delight impatient customers with a dazzling array of products, excellent service levels and low prices) is critical for business success across a variety of industries. Competing on product availability has become a key strategic weapon. This requires new data-driven methods rather than approaches that rely on distributional assumptions (which are the mainstay in current practice). There are five aspects that make the data complex. First, having a limited amount of purchase history makes it difficult to have a complete picture of the already uncertain consumer; thus, the parameter uncertainty in the model has to be accounted for, in addition to data uncertainty. Second, items are bought in sets, or act as substitutes, and so a tractable multi-variate model is needed that can be built with limited amount of data. Third, there are temporal aspects of purchase behavior, with auto-correlations and lags, due to consumer trends. Fourth, not every item is bought in every period, nor are they bought in fixed sums; this sporadic (intermittent) demand modeling is crucial as well, in addition to correlations between the time between purchases and purchase quantities. Finally, due to seasonality and life cycle phases, the purchase behavior is non-stationary. The investigators will use a framework that accounts for parameter uncertainty, a multi-variate representation that is at once sparse and general, and develop fast algorithms starting from available limited data by making the least amount of assumptions on the data generation process. Prior work was very successful in dealing with the single-product, limited history situation with no intermittent demand, and early testing shows very good performance with intermittent data. Beyond research, the investigators also plan to provide tools for educators so that these can be made part of the next generation of teaching tools for senior undergraduates and Masters students (including technical MBAs).