In　recent　years,　quantitative　investment　has　been　a　hot　spot　in　the　development　of　the　financial　market.　Quantitative　stock　selection　is　the　most　crucial　part　of　quantitative　investment.　It　is　of　great　significance　to　study　how　to　select　high-quality　stocks　from　thousands　of　stocks,　bring　them　into　the　stock　pool　and　allocate　assets.　Various　machine　learning　and　deep　learning　algorithms　have　been　used　in　this　research.　This　paper　proposes　a　new　stock　selection　strategy　for　multi-factor　anomaly　detection　based　on　variational　auto-encoder.　First,　we　select　factors　from　three　aspects:　fundamental,　technical,　and　capitalization.　Then,　unsupervised　anomaly　detection　is　performed　on　the　multivariate　time　series　data　based　on　variational　auto-encoder　to　obtain　the　anomaly　scores　of　the　factors,　and　get　the　abnormal　result　by　comparing　it　with　the　threshold.　Finally,　the　abnormal　results　are　used　to　select　stocks　combined　with　the　trend　of　the　selected　stocks.　We　apply　the　model　to　four　groups　of　stocks　belonging　to　SCI300,　SSE　50,　SZSI　and　CSI500　respectively,　and　evaluate　the　performance　compared　with　the　Buy&Hold　strategy,　Traditional　multi-factor　model　stock　selection　strategy,　AdaBoost　machine　learning　stock　selection　strategy.　The　experimental　results　show　that　the　model　can　identify　＇good＇　stocks　from　the　sample,　and　the　performance　of　the　selected　portfolio　is　better　than　the　benchmarks　test.　©　2022　IEEE.