I gather whatever set of projections I can; merge them if more than one. The only projected stats I use are IP, ERA, and saves. Only use saves to tweak the ERA; I might lower it by as much as 0.2 for a guy who is supposed to be among the saves leaders.
I use a formula like IP * (6 minus ERA) ^ 2
The square term captures how an elite ERA is more valuable; a run saved in a close game. I've also used a cube function like (7 minus ERA) ^ 3.
.