Dictionary compression, combined with a form of entropy coding can work. As shown in the article the exponents often repeat themselves, and the low order bits of the mantissa are often zero.
Advanced context modeling compression algorithms can work even better than the scheme presented in the article. These are like mini-AIs, and they can recognize patterns more complex than just a string of unchanged bits, and with the appropriate entropy coder, give excellent compression. Or course the processing cost is huge compare to a few bit manipulations, and wouldn't make much sense unless you need extreme compression.