Improved risk tail bounds for on-line algorithms