
Perishable inventory management remains a critical challenge for online grocery retailers, where mismatches between demand and shelf life can lead to substantial financial losses and food waste. This article describes the development and deployment of an integrated, data-driven system for managing near-expiry ("red-line") inventory across a nationwide network of microfulfillment centers operated by Meituan's Xiaoxiang Supermarket. The system jointly optimizes nighttime inventory rebalancing and daytime clearance pricing using a closed-loop learning-and-optimization framework. It combines continuous causal forest to estimate price elasticities, a linearized model for inventory rebalancing, and a Markov decision process for dynamic pricing. Across the full operational network, spoilage rates are about 20% lower and revenue from red-line inventory about 14% higher since deployment, with an annual financial benefit and an annual reduction in the value of food waste each estimated on the order of tens of millions of CNY. This case shows how analytics can be embedded into large-scale retail operations to improve both economic and sustainability outcomes.
One of the world's largest manufacturers relied on dozens of production planners using spreadsheets to coordinate production across multiple continents, hundreds of suppliers, and dozens of product families. The process absorbed contradictions through politics and manual overrides, producing plans that were feasible only because no single person ever saw all the constraints at once. The core optimization problem was textbooksimple on paper but resisted solution for years because the organization was not ready to trust what the math had to say. Mixed-integer linear programming was deployed incrementally, with capabilities chosen for what they let planners see rather than for raw algorithmic power: surfacing the smallest set of mutually inconsistent commitments behind an infeasibility, attaching a price to constraints that had been treated as absolutes, presenting a menu of near-optimal plans rather than a single prescription, and encoding planner priorities in a clear order of precedence. A natural-language interface and a Monte Carlo robustness layer were also built on top of this system; both warrant their own separate treatment, which we leave for future work. Over a two-year period, the optimization-driven production planning system freed up production capacity that the prior process had been leaving on the table, unlocking more than 40,000 additional finished units per year and approaching $400 million in realized profit improvement across one of the world's largest supply chains. Planning cycles that once consumed weeks ran in hours, and disruptions were met with structured scenario analysis rather than emergency meetings. The key to adoption was not algorithmic sophistication but tooling that made the model's reasoning legible to planners, preserving their agency and enabling trust-based organizational change.
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher approval, unless otherwise noted. For more information, contact permissions@informs.org. The Publisher does not warrant or guarantee the article's accuracy, completeness, merchantability, fitness inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or support
Even with advances in state-of-the-art hardware and software, large (mixed-integer) nonlinear programs can require hours, or even days, of runtime and are not guaranteed to yield an optimal (or near-optimal, or any!) solution. Nonetheless, with progress in software designed from its inception to handle nonlinear optimization models, such as IPopt, Knitro and BARON, and that which has evolved into incorporating algorithms that handle nonlinearities, such as those contained in CPLEX and Gurobi, the past decade has seen performance improvements from the determination of local optimality on models containing a few thousand variables and constraints to proof of global optimality on instances that are two orders of magnitude larger. Assuming basic knowledge of (mixed-integer) nonlinear programming formulation and a conceptual understanding of the algorithms applicable to these models, this paper provides a tutorial on efficiently solving large-scale nonlinear programs, which includes: (i) guidelines for careful formulation, (ii) the use of initial solutions, (iii) the provision of bounds, (iv) the principles of numerical stability, and (v) the use of the solvers to improve performance. Although applicable to all domains, we draw on three practical examples, one each from civil, mechanical, and electrical engineering.
Perishable inventory management remains a critical challenge for online grocery retailers, where mismatches between demand and shelf life can lead to substantial financial losses and food waste. This article describes the development and deployment of an integrated, data-driven system for managing near-expiry (“red-line”) inventory across a nationwide network of microfulfillment centers operated by Meituan’s Xiaoxiang Supermarket. The system jointly optimizes nighttime inventory rebalancing and daytime clearance pricing using a closed-loop learning-and-optimization framework. It combines continuous causal forest to estimate price elasticities, a linearized model for inventory rebalancing, and a Markov decision process for dynamic pricing. Across the full operational network, spoilage rates are about 20% lower and revenue from red-line inventory about 14% higher since deployment, with an annual financial benefit and an annual reduction in the value of food waste each estimated on the order of tens of millions of CNY. This case shows how analytics can be embedded into large-scale retail operations to improve both economic and sustainability outcomes. History: This paper has been accepted for the INFORMS Journal on Applied Analytics Special Issue—2025 Daniel H. Wagner Prize for Excellence in the Practice of Advanced Analytics and Operations Research.