Performance　of　the　case-based　reasoning　(CBR)　prediction　model　is　directly　affected　by　feature　weight　allocation　in　the　retrieval　process.　A　method　based　on　selfish　herd　optimizer-simulated　annealing　(SHO-SA)　algorithm　was　proposed　to　calculate　the　feature　weights.　The　root　mean　square　error　(RMSE)　of　the　CBR　prediction　model　was　first　defined　as　the　fitness　function　in　the　SHO　algorithm　and　the　SA　algorithm　to　evaluate　the　rationality　of　the　weight　distribution.　Then,　the　weights　distribution　with　the　minimum　RMSE　in　the　population　were　obtained　through　the　herd　movement,　predation　and　recovery　steps　of　the　SHO.　Finally,　SA　algorithm　was　employed　to　search　randomly　based　on　the　above　weight,　and　an　approximate　optimal　solution　of　the　feature　weights　was　obtained.　Performance　evaluation　was　carried　out　by　five　benchmark　regression　datasets　from　University　of　California　Irvine　(UCI)　datasets.　Results　show　that　compared　with　other　typical　optimization　methods,　the　proposed　method　can　improve　the　accuracy　of　the　CBR　prediction　model　significantly.　Meanwhile,　it　illustrates　that　the　matter　of　SHO　suffering　from　local　minima　can　be　mended　by　SA　algorithm.　©　2022,　Editorial　Department　of　Journal　of　Beijing　University　of　Technology.　All　right　reserved.