If your offset is a million digits long you need to be sending more than a million digits of data to make it worthwhile, but the chance of your chosen million digits being available in that space is effectively zero. :)
Regardless of the actual distribution of values in pi, it seems to me that the best case scenario can't compress all data on average when you include the offset and length, because of the pigeonhole principle. Same as any compression method, except practically worse/useless because pi isn't constructed to have the values that occur in practical situations come first.