I'll have a go at the answer, though I do not have a concrete program for you to try. In one Eelco Visser blog entry you can see a screen shot containing essentially the lambda calculus encoded in Stratego (I'm sure this code is available elsewhere). He presents an interpreter for the lambda calculus in Stratego. Presumably, this is sufficient to show that Stratego is Turing-complete.
If you are interested in a more specific notion of expressiveness, then I suggest that you rephrase your question.
In terms of practical usage, I found Stratego's main limitation that it operated over only one data structure, namely, the tree being transformed. Using other data structures such as a table of symbols was unnatural (at least in Stratego from 5+ years ago). Even though Stratego is ideal for transformations on a single tree, performing operations on two trees simultaneously is not easy.