It's calculated bet-by-bet.  And it's the number of bets that have been won expressed as a percentage of the number that you would expect to be won.

So if you play a 10% chance-of-winning game 2000 times and win 200 times (as expected) you get luck=100%.  If you only win 180 times, the luck is 90%.

The problem is when people play the 0.0001% game (one in a million chance).  Each time they lose (which is almost every time), the luck drops a bit.  Then when they win, the luck jumps right up.  It has been won twice already (see milestone list) in less than 2 million tries, so the overall luck is well over 100%, even though the bets themselves are mostly insignificant, being all single-satoshi bets.

If anyone can think of a better way of calculating "how lucky" the players have been (individually, and as a whole) I'd love to hear it!