Hedging Predictions for Operational Decision Making

在过去的几年里，随着数据的日益可用性，使用机器学习（和统计）技术进行预测的研究出现了爆炸式增长。然而，很少关注如何有效地使用这些预测来做出有效的操作决策，其中预测与需求（处理时间，需求数量）相关。特别是，随机建模文献几乎只关注预测没有错误或没有预测可用的情况（决策仅基于分布信息）。本提案从两个方向解决了文献中的这一空白。第一个重点是量化“错误预测的代价”，即天真地使用预测值代替真实值的成本，与只使用分布信息（没有预测）的成本进行比较。当错误预测的代价很高时，设计操作决策策略，从嘈杂的预测中提取有用的信息，同时对冲预测错误的潜在负面影响，就变得有趣了。这些政策的设计是本建议的第二个重点。为了深入了解这些问题，我们将考虑两个最流行的随机模型。第一种，排队模型，将在预测处理时间可用时进行研究。特别令人感兴趣的是扩大系统大小（服务器数量）对错误预测代价的影响，以及有效服务器调度和作业路由策略的设计。我们认为较大的系统对预测误差不太敏感。第二种是库存模型，将在需求预测可用时进行研究。我们将研究预测精度的影响和预测误差对所考虑的成本函数影响的敏感性，特别是建议确定安全库存以对冲预测误差的方法。对于这两种模型，可实现性将是一个关键问题。最优策略可能会变得相当复杂，因此确定简单的、接近最优的策略将是一个重要的考虑因素，这将使这项工作能够转化为实践。除了解决学术文献中的空白之外，这项工作有望产生实际影响。我们已经参与了血液制品库存管理项目，并且已经看到对某些产品的需求可以比其他产品更准确地预测，因此我们的结果应该为该领域有效的库存管理实践提供信息。医疗保健、零售销售、制造业和信息系统是数据丰富的领域的其他例子，与需求相关的潜在数量预测是可能的（或已经在执行）。因此，他们都是拟议研究的潜在受益者。

Over the past several years, with the growing availability of data, there has been an explosion in research on using machine learning (and statistical) techniques to make predictions. However, less attention has been paid to how to effectively use these predictions to make effective operational decisions, where the predictions are related to demand (processing times, demand quantities). In particular, the stochastic modelling literature is almost exclusively concerned with scenarios where predictions are error-free or where no predictions are available (decisions are based only on distributional information). This proposal addresses this gap in the literature in two directions. The first major focus is on quantifying the "price of misprediction", the cost of naively using predicted values in place of true values, measured against the cost of using only distributional information (no predictions). When the price of misprediction is high, it then becomes of interest to design operational decision policies that extract useful information from the noisy predictions while hedging against the potential negative impacts of prediction errors. The design of such policies is the second major focus of this proposal. To gain insight into these issues, two of the most popular stochastic models will be considered. The first, queueing models, will be studied when predictions of processing times are available. Of particular interest will be the effect of scaling up the size of the system (number of servers) on the price of misprediction and the design of effective server scheduling and job routing strategies. We believe that larger systems are less sensitive to prediction errors. The second, inventory models, will be studied when demand predictions are available. We will study the impact of the prediction accuracy and sensitivity of the impact of prediction error to the cost function being considered, in particular suggesting approaches to determine safety stocks to hedge against prediction errors. For both of these models, implementability will be a key concern. Optimal policies may become quite complicated, so the identification of simple, near optimal policies will be an important consideration that will allow this work to be translated to practice. Beyond addressing gaps in the academic literature, this work is expected to have practical impact. We are already involved in projects for inventory management for blood products, and have seen that demand for some products can be more accurately predicted than others, so our results should inform effective inventory management practices in this domain. Healthcare, retail sales, manufacturing and information systems are other examples of areas that are data-rich and predictions of underlying quantities related to demand are possible (or are already being performed). As such, they are all potential beneficiaries of the proposed research.