The paper touches on this. You could use a number which will become known on a future date. They propose the last 3 digits of a state run lottery.
Anything with enough randomness to cover your payout ratio would work. You could imagine taking a hash of your instance identifer concatenated with the NASDAQ final for the end date if you really don’t want to trust anyone.
And how do we all agree on the NASDAQ final for some end date now and in the past? Maybe the oracle cheated. So in any case we need multiple machines who we hope don't collude. How would we prove the NASDAQ final price was not manipulated ever so slightly by a colluding party? And so on.
Anything with enough randomness to cover your payout ratio would work. You could imagine taking a hash of your instance identifer concatenated with the NASDAQ final for the end date if you really don’t want to trust anyone.