The　echo　state　networks　(ESNs)　have　been　applied　in　many　applications.　For　an　ESN,　its　training　error　and　network　size　are　closely　related　with　output　weight　matrix.　In　this　paper,　the　coordinate　descent　method　based　ESN　(CD-ESN　for　short)　is　designed　to　deal　with　the　relationship　between　training　error　and　network　size.　In　CD-ESN,　the　l1　regularization　is　used　to　penalty　the　non-important　values　of　output　weight　into　0.　Moreover,　the　coordinate　descent　method　is　used　to　update　the　output　weights　of　ESN.　Experimental　results　imply　that　the　proposed　CD-ESN　has　better　prediction　accuracy　and　more　sparse　network　topology　than　original　ESN.　©　Springer　Nature　Singapore　Pte　Ltd.　2020.