In　intelligent　transportation　system　(ITS),　traffic　flow　prediction　can　provide　data　support　for　route　planning,　traffic　management　and　public　safety.　Prediction　algorithms　based　on　machine　learning　or　deep　learning　usually　need　a　large　number　of　unabridged　historical　data　to　conduct　parameter　training.　However,　data　will　be　missing　and　abnormal　in　practice,　which　will　affect　the　accuracy　of　prediction.　In　this　paper,　we　propose　the　stream　tensor　analysis　(STA)　algorithm　for　traffic　flow　prediction.　First,　dynamic　tensor　stream　of　four　dimensions　of　space,　time,　day　and　week　is　constructed　to　better　mine　the　multi-mode　correlation　between　traffic　flow　data.　Second,　the　first　few　columns　with　the　largest　norm　are　selected　to　update　the　projection　matrix　by　the　tracking　projection　matrix　algorithm.　The　experimental　results　show　that　the　STA　algorithm　has　low　complexity,　and　also　achieve　good　prediction　performance　in　random　missing　and　extreme　missing　patterns.　©　2021　ACM.