Description: In this article, we introduce a novel machine learning-based model that analyzes and predicts policyholder behaviors at the individual policy level. Policyholder embedding representations are learned using a Recurrent Neural Network (RNN) that takes into account the sequence of historical transactions associated with a policy. Additionally, a Neural Network with a multi-head architecture is trained to predict the policyholder's next action based on the policyholder embedding and other policy attributes. The resulting model offers several advantages over traditional experience studies and assumption-setting methods, including the ability to set policy-level assumptions that improve reserve accuracy, predict multiple behaviors simultaneously while capturing the interrelationships between them, and generate policyholder embedding representations that are useful for segmentation.