Source-linked AI summary
Data Analytics in Operations Management: A Review
Velibor V. Mišić, Georgia Perakis
TL;DR
Operations management has traditionally relied on strategic models, while growing data availability and advances in machine learning and optimization motivate broader use of analytics. This review synthesizes representative applications across supply chain, revenue management, and healthcare operations, and discusses future methodological directions. The reviewed applications include predictive and optimization-based methods with reported operational and revenue improvements, within a deliberately limited scope.
Problem
Operations management needs approaches that use increasingly abundant, granular data alongside advances in machine learning and optimization rather than relying primarily on traditional strategic models.
Method
The paper reviews representative applications of data analytics across supply chain management, revenue management, and healthcare operations, then discusses future methodological directions.
Results
The reviewed applications include data-driven inventory and pricing methods, a pricing pilot with about 10% higher revenues, and hospital scheduling that saves approximately $2.2 million per year.
Takeaways & Limitations
Analytics enables large-scale data to be used for complex operations-management decision-making, while causal inference, interpretability, and small-data methods are important future directions.
Takeaways & Limitations
Because of anniversary-issue page limits, the review is brief and covers only a small set of representative examples in each application area.
Abstract
from arXiv · showhide
Research in operations management has traditionally focused on models for understanding, mostly at a strategic level, how firms should operate. Spurred by the growing availability of data and recent advances in machine learning and optimization methodologies, there has been an increasing application of data analytics to problems in operations management. In this paper, we review recent applications of data analytics to operations management, in three major areas -- supply chain management, revenue management and healthcare operations -- and highlight some exciting directions for the future.
1 Introduction
Operations management research is shifting from primarily model-based strategic analysis toward data analytics as data availability and methodological capabilities expand. The review surveys representative analytics applications while acknowledging that its page-limited coverage is necessarily brief.
- Earlier operations management research mainly used theory-based models to generate strategic insights about how firms should operate.
- The field is shifting as richer, more granular data and advances in machine learning and optimization expand analytical possibilities.
- Analytics uses data to create models that support value-creating decisions across supply chain, revenue management, and healthcare operations.
- The review covers only a small set of representative examples because anniversary-issue page limits make it necessarily brief.
2 Analytics in Supply Chain Management
Analytics is being applied across supply-chain location, omnichannel, and inventory decisions by combining predictive models with optimization and decision rules. The reviewed work includes applications to truck positioning, fulfillment, service regions, contextual inventory, and dynamic procurement.
- Location and omnichannel operations: Omnichannel integrates e-commerce with brick-and-mortar stores, enabling cross-channel fulfillment and creating location and inventory decisions.
- Location and omnichannel operations: A random forest predicts demand by location and time to inform where and when a buy-online, pickup-in-store retailer positions delivery trucks.
- Location and omnichannel operations: A fulfillment heuristic minimizes immediate outbound shipping cost plus estimated future cost, capturing 36% of the opportunity gap in industry-data experiments.
- Inventory management: Inventory analytics can map demand features directly to order quantities through empirical risk minimization or conditional-demand modeling with kernel regression.
- Inventory management: 6-15% higher costs result when dynamic procurement ignores covariate information compared with the proposed predictive stochastic-programming approach.
- Inventory management: Data-driven demand estimation supports decisions for new products by identifying comparable products to predict demand.
3 Analytics in Revenue Management
Analytics is being applied across revenue management to model customer choice, optimize assortments, set prices and promotions, and personalize offerings. The reviewed studies combine predictive models with optimization or learning methods and report theoretical guarantees, approximation results, or revenue improvements in practical settings.
- Scope: Revenue-management analytics spans choice modeling and assortment optimization, pricing and promotion planning, and personalized revenue management.The reviewed applications address both what products to offer and how to price, promote, bundle, or personalize them.
- Choice modeling and assortment optimization: Ranking-based choice models can represent broad classes of customer behavior, while Markov chain models approximate true choice probabilities with less than 3% average maximum relative error.The ranking-based approach represents random-utility choice models; the Markov chain approach is exact for generalized attraction models under the stated conditions.
- Pricing and promotion planning: A machine-learning and integer-optimization pricing strategy increased Rue La La’s revenues by about 10% in a live pilot.Demand was predicted with a bagged regression tree, and prices were selected through integer optimization.
- Pricing and promotion planning: Promotion-planning methods exploit demand structure or optimization approximations, with reported potential profit improvements of 2-9% for promotion-vehicle scheduling.The bounded memory peak-end model enables compact dynamic programming and a PTAS for the multi-item problem, while promotion-vehicle methods compare greedy and mixed-integer approximations.
- Pricing and promotion planning: Analytics-based omnichannel pricing policies estimated a 13.7% increase in clearance-period revenue at a large US retailer.The policies use inventory partitions and solve deterministic or robust mixed-integer optimization problems with business-rule constraints.
- Personalized revenue management: Personalized pricing, assortment, and bundle decisions use customer features, inventory-aware heuristics, and learning methods, with reported revenue improvements of over 3% and 2-7%.These studies include guarantees relative to oracle policies, feature-dependent pricing algorithms, inventory balancing, and personalized bundle recommendations.
4 Analytics in Healthcare Operations
Analytics methods are being applied across healthcare policy, hospital operations, and patient-level medical decisions. These applications combine predictive modeling, machine learning, and optimization to address efficiency, fairness, staffing, scheduling, treatment, and dosing.
- Healthcare operations research spans policy-level, hospital-level, and patient-level problems.
- Policy-level problems: 8% increase in aggregate quality-adjusted life years was achieved while satisfying the same fairness criteria as the existing kidney allocation policy.The policy combines linear optimization with regression-based scoring of patient–kidney matches.
- Policy-level problems: Mixed-integer optimization selected chemotherapy combinations to maximize predicted median survival subject to predicted toxicity limits.Simulation and similarity-based evaluation both suggested improved efficacy for regimens tested in Phase III trials.
- Hospital-level problems: $2.2 million per year in anesthesiologist overtime was saved after implementing robust optimization for surgery staffing and scheduling.The methodology accounts for uncertainty in surgery duration and is in use at Ronald Reagan Medical Center.
- Hospital-level and patient-level problems: Hospital analytics also addressed surgical-flow congestion, emergency-department wait-time prediction, and patient-specific dosing.Reported approaches include optimization, LASSO combined with queue-derived predictors, and a LASSO-based bandit algorithm.
- Patient-level problems: A k-nearest-neighbors approach was used to recommend personalized diabetes drug regimens aimed at minimizing HbA1C.
5 Conclusions and Future Directions
The review identifies methodological directions for making analytics more decision-valid, interpretable, effective in small-data settings, and directly aligned with prescriptive objectives. It also illustrates these directions through causal inference, personalized modeling, and decision-focused estimation.
- Conclusions: The review concludes that machine learning and optimization enable large-scale data to support complex operations-management decisions.
- Causal inference: Causal inference is needed because decisions often enter machine-learning models as variables, making causal effects more relevant than prediction alone.
- Causal inference: Confounding can make regression-based data-driven pricing highly suboptimal, motivating tests of whether price prescriptions are optimal.
- Interpretability: Interpretable models can reveal prediction structure and help human decision makers accept recommendations, especially in medicine.Legislation such as the GDPR also creates requirements for explanations of algorithmic decisions.
- Interpretability: MinOP uses interpretable machine learning on exact stochastic dynamic-program solutions to expose the structure of optimal policies.
- “Small data” methods: In the “small data” regime, many observations coexist with many parameters, making standard sample-average approximation suboptimal.Empirical Bayes and regularization approaches provide theoretical guarantees as the number of uncertain parameters grows.
- “Small data” methods: Tensor completion and related methods extend small-data analytics to sparse, noisy applications such as cancer blood-test design.
- New approaches to “predict-then-optimize”: Decision-focused estimation replaces predictive-loss minimization with a loss related to downstream objective value, improving decisions in shortest path, assignment, and portfolio problems.