All Publications
In many decision processes, a decision maker or planner must review and optionally adjust the recommendations that are generated by a decision support system (DSS). When the DSS is well-tuned to its task, adjustments by a planner can be rare and may even degrade the DSS’s performance. Targeted automation could address these inefficiencies by predicting whether a planner will adjust a recommendation and improve the performance of the system. The remaining recommendations can be automated. However, as more recommendations are automated, fewer will receive planner input. This may starve the prediction model of the observations it needs for retraining. To maintain predictive performance, we must therefore address the loss that automation imposes on the model’s ability to learn from a planner’s decisions over time. Using 4 years of procurement ordering data from our research partner, a large materials handling equipment manufacturer, we develop and train a series of machine learning classifiers that predict individual instances in which a planner will improve a DSS-generated procurement order decision. We mitigate the performance erosion that automation engenders by structuring the selection of the model’s classification threshold similar to a newsvendor problem, accounting for the value of learning and balancing the costs and benefits of under or over automating. In our setting, this approach automates around 84% of all DSS recommendations while retaining three times more planner improvements than random automation. The models maintain their predictive performance over time, despite losing automated outcomes for retraining and substantial dataset shift. Our research contributes to a broader debate on the allocation of decision authority between humans and algorithms, and creates a framework for targeted automation in an operational setting that balances the net benefits of automation versus the long-term benefits of algorithmic learning.
When binary classification models are wrong, managers face misclassification costs. Although false positive outcomes imply unnecessary mitigation efforts, false negative outcomes imply overlooking the class of interest. Humans calibrate these ai models supporting operational systems by adjusting the decision threshold that translates prediction probability into either class. Results of our controlled laboratory experiment show that, despite all relevant information being available, decision makers systematically deviate from the optimal cost-efficient threshold. We observe a significant interaction effect of class and cost imbalance on this deviation, which increases in high-stakes settings where more extreme thresholds are optimal. When unit costs are different, we find that participants anchor on the threshold where expected misclassification costs for false alarms and missed hits are equal, whereas mean anchoring cannot explain the pull-to-center behavior sufficiently. Surprisingly, we confirm that this impulse balance equilibrium also serves as attractive anchor in our setting, where decisions are made ex ante without loss aversion. To debias decision makers, simulated responses with behavior-aware costs show that subjects are nudged to make choices closer to the optimum. Managers should be aware of this boundedly rational behavior and complementary debiasing techniques, as sub-optimal threshold setting results in 53% higher misclassification costs, on average.
Understanding how inventory policies are implemented in practice remains a fundamental challenge in operations management because replenishment rules are rarely shared with others. However, different inventory policies leave identifiable signatures in transactional order data, creating an opportunity to infer otherwise hidden replenishment structures. This study develops a two-stage data-driven framework for identifying inventory policies directly from observed order behavior. In the first stage, semi-synthetic data, machine-learning classification, and expert evaluation are used to develop and validate a robust policy-identification model. In the second stage, the selected model is applied to a large-scale order dataset provided by a global travel-retail wholesaler, comprising 128,834 order time series across a diverse set of retailers and product categories. The results reveal only partial alignment between observed replenishment behavior and classical inventory-theory expectations. While policy selection varies systematically with product characteristics, substantial heterogeneity exists across retailers, and identical products are frequently managed using different replenishment rules. Retailer-level factors explain more variation in inferred policy behavior than product characteristics alone, highlighting the importance of organizational context in inventory-policy implementation. The study contributes a scalable methodology for inventory-policy identification and provides large-scale empirical evidence on how inventory policies are implemented in practice. More broadly, it demonstrates how the digital traces embedded in transactional order data can be used to uncover otherwise unobservable operational decision structures.
Internet-of-Things-enabled systems that monitor usage and inventory are the latest technological advancement in demand forecasting and inventory control. Unlike traditional systems that record sales via cash registers or RFID technology at the point-of-sale, these novel systems can track product usage via smart, connected devices at the point-of-consumption, i.e., directly at the end user. This usage data promises to be a valuable basis for smart, automated replenishment services. We study such a service in the context of commercial coffee machines through collaboration with a large manufacturer in the coffee industry. Our data set contains information on more than 75 million drinks recorded since late 2017 by nearly 6,500 IoT-enabled coffee machines for commercial customers such as office kitchens, restaurants, and gas stations. The nature of the problem and data at the point-of-consumption warrants the development of synergetic models for demand forecasting, inventory control, and correction of inventory record inaccuracy. The resulting models are distinct from the state-of-the-art approach at the point-of-sale as they are uniquely integrated and involve an alternative strategy to mitigate inventory record inaccuracies. Overall, we contrast different approaches to manage smart replenishment systems, test their forecasting, inventory control, and inaccuracy correction performance, and pave the path to implementation in the field. Our findings suggest important implications for manufacturers who wish to engage in direct relationships with the end users of their products.
Firms are exposed to varying levels of supply chain risk and engage in efforts to resolve such risk. This paper examines how disclosures of supply chain risk and resolution during earnings calls affect firms? stock returns. Using natural language processing, we develop measures of supply chain risk and resolution from quarterly earnings call transcripts for a total of 129,981 firm-quarter observations between 2008 and 2019. We find that higher levels of supply chain risk are associated with lower stock returns around earnings calls, while disclosures of supply chain risk resolution attenuate these negative effects. In particular, stock returns of firms in the highest supply chain risk quintile are 1.07% lower compared to the stock returns of firms in the lowest quintile, and regression analyses indicate that a one-standard-deviation increase in supply chain risk is associated with a 0.56% decline in stock returns. The stock returns of firms in the highest supply chain risk resolution quintile are 0.12% higher compared to the stock returns of firms in the lowest quintile. A one-standard-deviation increase in resolution increases stock returns by 0.08%, and to 0.29% for the subsample of observations where the resolution measure is positive. Exploratory analyses indicate that the effect of supply chain risk on stock returns is significantly greater for smaller firms than for larger firms. In addition, when there is evidence that larger firms? resolution-related statements are mere rhetoric, the effect of resolution on stock returns is diminished.
Manufacturing firms face complex after-sales challenges, including spare part shortages. While additive manufacturing (AM) offers a solution by minimizing costs and complexity, not all firms adopt AM equally, and research on differences in AM adoption in the context of spare part shortages is surprisingly scarce. To close this knowledge gap, we apply the awareness-motivation-capability (AMC) perspective. Our comparative case study of AM applications in 17 firms identifies three approaches how firms adopt AM—the corrective, preventive, and anticipatory approach. We find that the specific configuration of contextual factors related to a spare part shortage determines the approach firms follow. Using the AMC perspective, we discover and explain why firms differ in adopting AM despite suitable spare part characteristics and similar contexts. Through uniquely analyzing spare part shortages, our study contributes to AM research by challenging the assumption that economic justification is the sole driver of AM adoption and instead revealing that it is a context-dependent process, with awareness and motivation serving as critical yet underexplored antecedents.
Many customers complain when informed that their order will not be fulfilled as originally confirmed, while other customers may be able to tolerate deviations. However, for suppliers, such complaints can be an early indicator of bad publicity, customer churn, and lost sales; and suppliers can prioritise orders to avoid these negative consequences. Ideally, they would know in advance if any order fulfilment change will trigger a customer complaint. To analyse how suppliers can predict these infrequent events in a business-to-business context, we leverage machine learning models on a large real-world dataset from a global semiconductor manufacturer. Our findings demonstrate that extreme gradient boosted trees effectively address the prediction problem. We explore the impact on model performance for different sampling approaches and cutoff values, as tuning the decision threshold is a meaningful calibration strategy before practical implementation. Our feature importance analysis provides evidence that high order fulfilment quality lowers complaint tendencies. Bridging the gap between advanced analytics and customer behaviour prediction, our research contributes to understanding the influence of subpar order fulfilment on customer satisfaction and offers insights into efficient order management despite disruptions. Our empirical study lays the groundwork for proactive supply chain operations when order fulfilment is at risk.
Mobile digital factories are production units equipped with direct digital manufacturing (DDM) technology that can be set-up at any location and moved to another place of demand after completing the production job. From a theoretical standpoint, this manufacturing approach combines the concepts of direct digital, on-site, and mobile manufacturing. While the operation of mobile digital factories is currently explored by companies in various industries, there is little theoretical understanding about how they deliver value in supply chains and what role DDM plays in that. Our qualitative case data from the construction industry?in which mobile factories fitted with different technologies are an established manufacturing concept?provide insights into the complex interrelations between a mobile digital factory?s value functionalities. 32 semi-structured interviews and secondary data reveal that logistics efficiency, responsiveness, geographical flexibility, and production time efficiency are the key value outcomes, and that DDM technologies are enabling their realization. Moreover, use cases emphasize that the project context influences the extent to which mobile digital factories are efficiently operated. Reflecting on the case study insights, we theorize that (i) supply chain orchestration enable the responsiveness and geographical flexibility and (ii) postponement and speculation is explanatory for logistics and production time efficiency. We conclude that mobile digital factories should be seen as a complement to traditional off-site supply chains that can balance flexibility and efficiency in a limited number of projects.
Demand uncertainty can lead to excess inventory holdings, capacity creation, emergency deliveries, and stock-outs. The costs of demand uncertainty may be directly borne by upstream suppliers, but can propagate downstream in the form of higher prices. To address these problems, we investigate a practical application of a fixed order commitment contract (FOCC) in which a manufacturer commits to a minimum fixed order quantity each period and receives a per unit price discount from the supplier for the commitment. We model a FOCC as a Stackelberg game in which the supplier offers a price discount anticipating the manufacturer’s response, and the manufacturer subsequently decides on the optimal commitment quantity. We show that a FOCC can smooth the orders received by the supplier, mitigating the negative consequences of demand uncertainty for the supplier, the manufacturer, and the supply chain. We extend the current literature by solving for an endogenous price discount instead of treating it as an exogenous value, and validate our model insights with our research partner, a large international materials handling equipment manufacturer. Using data on 863 parts, we evaluate the relationships between the model parameters, contract parameters, and the contract effectiveness, and show the conditions under which the FOCC generates greater cost savings for both the manufacturer and supplier. Our results help operations managers better understand how to obtain the optimal contract parameters for a FOCC and the circumstances under which such a contract is most beneficial for the company and its supply chain.
In many real-world situations, multiple humans are involved in decision-making when interacting with machine recommendations. We investigated a setting where an artificial intelligence system creates demand forecasts that a human planner can either accept or revise, and a supervisor then makes the final decision about which forecast to select. We designed and conducted two experimental studies to understand decision-making by a supervisor. First, we provided the improvement probabilities of adjustments at an aggregated level and found evidence for overoptimism bias and mean anchoring. Second, we provided decomposed guidance based on two adjustment attributes, direction and magnitude, to investigate the role of salience based on the distance between the improvement probabilities and level of detail in guidance effectiveness. We found no significant difference in using less and more salient guidance provided that the detail level was fixed. However, revealing more details when the guidance was more salient increased the use of guidance.
Recent advances in process mining technology have extended its applicability beyond traditional domains such as healthcare, finance and manufacturing, making it increasingly relevant for addressing problems in supply chain management. In our research, we explore the integration of process mining techniques within the domain of supply chain management, focusing on uncovering inefficiencies, ensuring compliance, and identifying opportunities for improvement. Therefore, we first review the technological advances in process mining relevant to supply chain management and outline six relevant approaches. We then combine the identified methodologies with expert interviews to derive and validate six specific use cases where process mining can significantly contribute to supply chain efficiency and resilience. The paper presents a detailed description of these use cases, demonstrating how process mining can provide actionable insights for a wide range of supply chain
problems. We discuss the implications of our findings for practitioners, who benefit from enhanced visibility and optimisation opportunities, and researchers, who are provided with a roadmap for exploration of this promising interdisciplinary field. To the best of our knowledge, this is the first work to explore potential use cases for process mining in a supply chain context, providing a comprehensive perspective on the potential benefits and challenges.
Many new ventures position themselves along the additive manufacturing (AM) value chain to benefit from the quickly maturing technology. Yet, their business models and sources of value creation are largely unidentified. We compile a unique dataset with 160 entrepreneurial AM-firms using a leading crowd-sourced database and organize the data in a card file system. We code the data along multiple dimensions and apply a Latent Class Analysis to identify unique segments of firms that focus on different complementary activities along the AM value chain. By using the NICE framework, we additionally reveal the value creation mechanisms of each identified class. We identify four unique segments of AM-firms that focus on different complementary activities: (i) hardware providers, (ii) software and data experts, (iii) full-service providers, and (iv) manufacturing orchestrators. While a lot of value creation across these segments is currently still driven by novelty and innovation, AM-firms also introduce lock-in and complementary products and services to capture value beyond production. In characterizing value chain structures, we outline how firms can position themselves in this emerging industry.
Slow-moving goods are common in many retail settings and occupy a vast part of retail shelves. Since stores sell these products irregularly and in small quantities, the replenishing distribution center may only place batched orders with manufacturers every few weeks. While order quantities are often fixed, the challenge for manufacturers facing such intermittent demand is to forecast the order timing. In this paper, we explore the value of Point-of-Sales (PoS) data to improve a food manufacturer’s order timing forecast for slow-moving goods. We propose an inventory modeling approach that uses the last order, PoS data from retail stores, and the expected lead time demand to estimate the retailer’s channel inventory. With this dynamic estimate, we can ‘nowcast’ the retailer’s inventory and predict his next order. To illustrate our methodology, we first conduct an experimental simulation and compare our results to a Croston variant and a moving average model. Next, we validate our approach with empirical data from a small German food manufacturer that serves a grocery retailer with a central distribution center and 53 hypermarkets. We find that, on average, our approach improves the accuracy of order-timing predictions by 10–20 percent points. We overcome a shrinkage-induced bias by incorporating an inventory correction factor. Our approach describes a new way of utilizing PoS data in multi-layered distribution networks and can complement established forecasting methods such as Croston. Particular applications arise when the order history is short (e.g., product launch) or represents a bad predictor for future demand (e.g., during COVID-19).
The emergence of digital technologies across all aspects of operations management has enabled shifts in decision making, shaping new operational dynamics and business opportunities. The associated scholarly discussions in information systems and operations management span digital manufacturing, the digitalization of operations management and supply chain management, platform outcomes, and economies of collaboration. For such changes to be successful, however, there is a need for organizations to go beyond the mere adoption of digital technologies. Instead, successful changes are transformational, delving into digital transformation endeavors, which in turn can enable operational improvements in organizational performance, lead to structural changes in operations processes, and may result in new business models being deployed. Our aim here, thus, is to provide an epistemic platform to advance our understanding of how such endeavors, including the adoption of digital technologies, business model innovations, and innovations in collaboration mechanisms and methods of operations improvement, can affect various aspects of operations management.
Complete and accurate data is an important enabler of effective supply chain decision making. Despite the increasing efforts to fully automate data collection processes using advanced sensors and scanners, human operators are still in charge of data entry tasks in most industries. Unfortunately, operators do not often comply with the standard operating procedures (SOPs) and do not always exhibit the consistency and commitment required to collect high-quality data. In fact, data collection is often perceived as a non-value-adding activity that increases workloads and lowers productivity. We aim to empirically study the extent to which compliance with SOPs for data collection is affected by some of the key factors. Using a large dataset obtained from a leading postal service provider in Australia, we find that an operator’s workload, fatigue, and related work experience directly impact the compliance levels. We also find that a company’s compliance reinforcement intervention to improve compliance behavior can moderate these impacts.
