Dual sourcing under non-stationary demand and partial observability

Hannah Yee*, Heletjé E. van Staden, Robert N. Boute

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review


We study dual sourcing under stochastic and non-stationary demand. The non-stationarity is modeled through Markov-modulated changes in the underlying demand distribution. The actual demand distribution is not observed directly, yet demand observations reveal partial information about it. We propose a policy where a pre-committed base order from the slow source is complemented with flexible short-term orders from both the fast and slow source. The pre-committed order is cheaper, while flexible orders can be adjusted to the actual inventory needs and the non-stationary demand. By formulating the problem as a partially observable Markov decision process, we show that the optimal flexible orders follow an adaptive dual base-stock policy when the lead time difference between both sources is one period. A numerical validation study reveals how flexible slow source orders reduce the share of expensive orders from the fast source compared to a conventional tailored base-surge policy. In addition, our policy's ability to adapt decisions to partial information allows for a more effective use of flexible orders. Our findings show the value of incorporating partial information to deal with the non-stationary demand and adding the flexible slow-sourcing option to create a more resilient replenishment policy.

Original languageEnglish (US)
Pages (from-to)94-110
Number of pages17
JournalEuropean Journal of Operational Research
Issue number1
StateAccepted/In press - 2023


  • Dual sourcing
  • Inventory
  • Non-stationary demand
  • Partially observable Markov decision process

ASJC Scopus subject areas

  • General Computer Science
  • Modeling and Simulation
  • Management Science and Operations Research
  • Information Systems and Management


Dive into the research topics of 'Dual sourcing under non-stationary demand and partial observability'. Together they form a unique fingerprint.

Cite this