On the use of case estimate and transactional payment data in neural networks for individual loss reserving