Absolutely. The better the coding exercise models the actual work you'd be doing in the job, the better of a predictor the test is.

Google found that a "work sample test" (when done well!) was able to account for 29% of the candidate's actual job performance.

(source: Work Rules! by Laszlo Bock)