The proposed selfish mining strategy in a nutshell is: when you find a block, keep it secret in the hope of finding a second one that will give you the leverage to start messing around. If someone else finds a block before you find your second, go ahead and publish.
Since by hypothesis you have a minority of the total computing power, usually someone else will find a block before you find your second.
But by the time you become aware of this, at least some other miners are also aware of it, whereas no others are yet aware of your block. With the 'first wins' tiebreaker, usually that means your block gets lost. So most of the time you lose out. So on average, selfish mining loses, at least as long as your computing power is small compared to the network total.
Is there something I'm missing?