In　a　complex　supply　chain,　product　demand　is　characterized　by　non-linearity　and　instability.　In　particular,　demand　trajectory　with　intermittent　or　abnormal　spikes　exists,　which　makes　it　difficult　for　suppliers　to　accurately　estimate　the　time-varying　demand　distribution　and　make　the　accurate　inventory　decision.　Hence,　a　deep　autoregressive　model　with　a　time　attention　mechanism　(Attention-DeepAR　model)　was　proposed　to　overcome　this　practical　issue.　In　addition,　a　separated　estimation　and　optimization　approach　was　provided　to　solve　a　newsvendor　problem　with　shifting　demand.　Temporal　features　were　extracted　from　historical　data　through　the　Attention-DeepAR　model　to　identify　the　long-term　and　short-term　trends　of　the　demand.　Accordingly,　the　time-varying　demand　distribution　was　accurately　fitted,　which　can　assist　in　making　precise　inventory　decisions.　In　addition,　the　rolling　window　design　was　proposed　to　introduce　new　demand　data　to　update　the　demand.　Finally,　the　validity　of　the　Attention-DeepAR　model　was　verified　through　the　Monte　Carlo　simulations　and　a　real　case.　Results　show　that　the　Attention-DeepAR　model　can　effectively　capture　the　temporal　correlation　between　the　regular　and　promotional　demand　values　under　volatile　demands,　improving　the　fitting　accuracy　of　the　time-varying　demand　distribution.　This　model　can　provide　precise　inventory　decisions　and　significantly　decrease　average　total　costs.　©　2023　School　of　Science,　IHU.　All　rights　reserved.