I was puzzled by that too. It seemed like the weakest thing in the paper. Other stuff, especially the part about applying contemporary risk models to historical data, was much more interesting.