in the past few years it worked quite nicely but that still is not bullet proof. and we had to refine our methodology further.
for example one candidate solved the task perfectly, but on the end of the "trial day" the candidate was unable to answer some of our questions regarding his code. so we think that he had access to the task details upfront, or someone else helped;
another example was with candidate that solved the task, answered our questions, but in the office was lacking everything that he presented to us in the "trial day". we suspect that the relative of the candidate helped here, as we found later that the relative has job in similar position in another company.
so the "trail day" is good way to test candidate, unless he has early access to the details of the task, or relative is helping with the task. it sorted a lot of candidates that were unable to follow the task details, or worse lacked team working skills