EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Adaptive Risk-Aware Decision-Focused Learning for Dynamic Inventory Optimization with Lead Times and Censored Demand

Jasper Moore, Mia Edwards, Katherine Kang

SOCIAL AND MANAGEMENT SCIENCES RESEARCH JOURNAL · 06.10.2026 · 4 (2) · 254-268

DOI: 10.66238/smsrj.v4i2.86 ↗

Kurz gefasst

Fachartikel in der Zeitschrift SOCIAL AND MANAGEMENT SCIENCES RESEARCH JOURNAL, online erschienen am 06.10.2026. Verfasst von Jasper Moore und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Decision-focused learning trains a demand forecaster on the cost of the decisions it induces, but in retail it has mostly been applied to single-period decisions, whereas replenishment is dynamic: stock carries over, orders arrive after a lead time and sales are censored when shelves empty. This paper proposes DynRA-DFL, an adaptive risk-aware decision-focused learning framework for periodic-review lost-sales systems. A global network emits monotone multi-horizon demand quantiles and a state-conditional risk weight; an order-up-to layer mixes the critical-fractile level with a closed-form CVaR-optimal level; and the network is trained through a differentiable simulator with deterministic or stochastic lead times, promotion-scaled penalties and a censoring-corrected mean–CVaR cost. An online risk calibration (ORC) rule updates a per-series safety margin from stock-out indicators and provably bounds the long-run stock-out frequency. On 909 M5 and 281 Stallion series, evaluated on observed sales and on reconstructed demand, risk-neutral dynamic training attains mean costs 11.2% lower than static single-period decision-focused learning and 5.9% lower than the same network used for predict-then-optimise, while adding CVaR only at the decision stage raises cost. DynRA-DFL trades off average inventory cost against selected tail-risk reductions: relative to the risk-neutral dynamic baseline it lowers the per-item tail cost by 10.4% at a 2.9% higher mean cost. Under a simulated demand surge, ORC raises the fill rate of the affected series for every policy, with cost changes ranging from −2.5% to +1.4%. We also quantify the bias of evaluating on censored sales.

Zitieren

APA

Moore, J., Edwards, M., & Kang, K. (2026). Adaptive Risk-Aware Decision-Focused Learning for Dynamic Inventory Optimization with Lead Times and Censored Demand. SOCIAL AND MANAGEMENT SCIENCES RESEARCH JOURNAL, 4(2), 254-268. https://doi.org/10.66238/smsrj.v4i2.86

Vancouver

Moore J, Edwards M, Kang K. Adaptive Risk-Aware Decision-Focused Learning for Dynamic Inventory Optimization with Lead Times and Censored Demand. SOCIAL AND MANAGEMENT SCIENCES RESEARCH JOURNAL. 2026;4(2):254-268. doi:10.66238/smsrj.v4i2.86

BibTeX
@article{Moore2026,
  author = {Moore, Jasper and Edwards, Mia and Kang, Katherine},
  title = {Adaptive Risk-Aware Decision-Focused Learning for Dynamic Inventory Optimization with Lead Times and Censored Demand},
  journal = {SOCIAL AND MANAGEMENT SCIENCES RESEARCH JOURNAL},
  year = {2026},
  volume = {4},
  number = {2},
  pages = {254--268},
  doi = {10.66238/smsrj.v4i2.86}
}
Kostenlos starten →