The mantissa of a double has 52 bits°, you need 50 bits to encode 15 decimal digits ( 2**10 =1024, so rule of thumb 10 bits for 3 decimals)

So the last two bits are used for error correction, since they don't suffice to encode another decimal digit.

Is it a good solution? Well I'd say appropriate for the era and not "for a dubious reason".

Are there better approaches? Sure, but they are harder to implement and would have been quite slow back in the days.

From my experience, >95% of the problems arise from calculating with currencies and the solution is obvious, calculate with integer cents and move the point only for output.

°) WP actually says 53 by using a spare bit redundancy , but tl;dr and this still fails to encode 16 decimals.