Supply Chain (CESIT)

However, dominant approaches in supply chain practices and research, including S&OE, still adopt a predict-then-optimize paradigm, where substantial effort is placed on improving forecast accuracy (e.g., demand, lead times, available capacities, etc.) with the expectation that better predictions yield better decisions. Recent works in predictive and prescriptive analytics such as (Sadana et al., 2024) or (Donti et al., 2021) have shown that this assumption is misleading and that the value of predictive models must be assessed through decision quality, not statistical accuracy. Improving prediction accuracy does not necessarily reduce decision regret, especially when operational costs are non-linear, constraints are binding, and uncertainties interact with decisions in complex ways (Bertsimas & Kallus, 2020; Elmachtoub & Grigas, 2022). This issue is particularly acute in S&OE whose purpose is precisely to continuously adjust the operational plan according to deviations from reality (delays, breakdowns, contingencies, urgent requests, etc.). 

In this context, alternative approaches must be considered. Recent advances in Contextual Optimization (CO) (Sadana et al., 2024) and Decision-Focused Learning (DFL) (Donti et al., 2021) offer a promising foundation for developing innovative models that could significantly enhance the effectiveness of S&OE processes. By integrating contextual information (e.g., current inventory levels, transportation constraints, production disruptions) into the optimization process itself, CO enables the generation of decisions that are robustly tailored to the operational context rather than to an abstract statistical forecast. Similarly, DFL couples Machine Learning (ML) and optimization in an end-to-end differentiable pipeline, ensuring that the predictive component learns representations that are explicitly optimized for downstream decision quality. In the volatile and constraint-rich environment of S&OE, where decisions must be revised continuously under uncertainty, these approaches remain under-explored while they could thus transform reactive execution into proactive, learning-driven adaptation, improving responsiveness, resilience, and economic performance across the short-term planning horizon.

While the contributions of these methods to the problem posed by S&OE are obvious, there remain numerous scientific obstacles, since S&OE decisions are inherently multi-stage, context-dependent, and time-sensitive, requiring sequential re-optimization as new information unfolds. However, to date, known studies on CO or DFL remain limited to simplified or single-period problems (e.g., contextual newsvendor, portfolio allocation). As a result, the potential of contextual, ML-integrated optimization to enhance short-term execution performance in global multi-echelon supply chains has not yet been rigorously established.

This research addresses this gap by developing CO and DFL models specifically tailored to S&OE, integrating ML and optimization in a unified framework. The objective is to shift S&OE from a forecast-centric approach toward a decision-centric, context-aware prescriptive paradigm that explicitly minimizes execution cost and regret under uncertainty, while ensuring service continuity and operational feasibility.

Research Questions: This work will address the following research questions:

• How can contextual information be formally modeled and integrated into S&OE decision-making framework to capture its influence on uncertainty, feasibility conditions, and costs in supply chains?
• In a context of uncertainty management in S&OE, what modeling strategies can effectively integrate ML and optimization to achieve high-quality continuous adaptive execution decisions?
• Under which structural and environmental conditions do context-aware, S&OE ML-integrated optimization models outperform forecast-driven and non-contextual prescriptive approaches in terms of regret and service–cost performance?

Methodology, Contributions and Dissemination:

In relation to the three research questions, the proposed methodology will follow three stages combining conceptual modelling, methodological development, theoretical analysis, and empirical validation.

Formalizing Context for Execution Decisions. This stage will establish a structured representation of context for S&OE, identifying the internal and external factors that materially influence execution outcomes and clarifying how they modify uncertainty, feasibility, and cost structures. 

Contribution #1: A context-driven S&OE decision-making framework. 
To be promoted through a scientific paper to be published in JOM, IJOPM, IJPDLM, DSS or equivalent. 

Developing Contextual Optimization Models. Models integrating ML and optimization will be designed for rolling execution decisions using two complementary paradigms:

• Sequential Learning and Optimization (SLO): learning produces context-conditioned inputs (e.g., predictive demand or lead-time distributions) that feed stochastic or robust execution models. This reflects current industrial integration practices and provides a benchmark for contextual prescriptive modelling.
• Integrated Learning and Optimization (ILO): predictive models are trained with decision-focused objectives so that learning directly improves execution outcomes rather than statistical accuracy. This enables decision-optimal learning in multi-stage, capacity-constrained execution settings.

Both approaches will be formulated for rolling, multi-stage execution, reflecting the continuous re-optimization inherent to S&OE. SLO enables context-aware integration within existing planning architectures by improving the quality of decision inputs, whereas ILO directly targets execution performance by minimizing regret rather than prediction error. This comparison will show when improving predictions is sufficient and when learning must be optimized for the decision itself. The models will be assessed for decision quality, regret, robustness to context shifts, and sensitivity to misspecification. Controlled computational experiments will benchmark contextual approaches against forecast-driven and non-contextual prescriptive baselines, yielding conditions under which SLO or ILO is preferable for execution decisions.

Contribution #2: A S&OE contextual optimization model based on SLO. 
To be promoted through a scientific paper to be published in MSOM, EJOR, IJPE, IJPR or equivalent. 

Contribution #3: A S&OE contextual optimization model based on ILO. 
To be promoted through a scientific paper to be published in MSOM, EJOR, IJPE, IJPR or equivalent. 

Empirical Validation and External Generalizability. The models will be tested on real datasets and case studies. Evaluation will focus on service performance, cost-to-serve, and robustness. This will establish the external validity and practical relevance of contextual optimization for S&OE. This part of the research will be carried out with the support of one or more industrial partners to contribute by providing us with access to one or more testing grounds. Several potential partnerships are already being explored with companies known to the doctoral thesis project leaders. 

Contribution #4: An Empirical Study based on previous models and concepts. 

To be promoted through a scientific paper to be published in POM, OMEGA, PPC, C&IE or equivalent. 

References

• Ban, G. & Rudin, C. (2019). The Big Data Newsvendor: Practical Insights from Machine Learning. Operations Research.
• Bertsimas, D. & Kallus, N. (2020). From Predictive to Prescriptive Analytics. Management Science.
•  Elmachtoub, A. N. & Grigas, P. (2022). Smart Predict-then-Optimize. Management Science.
• Donti, P., Amos, B. & Kolter, Z. (2021). Task-based Learning for Stochastic Optimization.
• Mišić, V. & Perakis, G. (2020). Data Analytics in Operations Management.
• Sadana, U., Chenreddy, A., Delage, E., Forel, A., Frejinger, E. & Vidal, T. (2024). A Survey of Contextual Optimization.