A0648
Title: Stochastic loss reserving with long short term memory
Authors: Yuning Zhang - The University of Sydney Business School (Australia) [presenting]
Boris Choy - University of Sydney (Australia)
Junbin Gao - The University of Sydney (Australia)
Abstract: A flexible mixture density network (MDN) approach for stochastic loss reserving in general insurance is proposed. To model the temporal sequences of claim losses presented in the run-off triangle, a special bi-directional Long Short Term Memory (2DLSTM) is employed. Unlike the original bi-directional Recurrent neural network (biRNN) and bi-directional LSTM, which train the model in both forward and backward time directions, the 2DLSTM uses input information from the top and left neighbours of the run-off triangle during the training procedure. This allows the proposed approach to capture both the accident period and development period dynamics in loss reserving.