The capability of predicting failures in Hard Disk Drives is a major objective of many data storage service providers. Such capability may avoid severe data losses, increasing the quality and reducing the costs of the service provided by such companies. Considering several recent works, Recurrent Neural Networks are shown to be the most promising methods for failure prediction on HDDs. Based on that results, in this work, we aim to evaluate the performance of several recurrent neural models in the HDDs failure prediction task. We conducted a comparative study between shallow and deep architectures and verified that deep models had the best results. Additionally, we evaluated the performance of all models when using three initialization procedures.