The multifaceted notion of time is gaining increasing interest in Information Retrieval (IR) systems. The popularity of microblogging services such as Facebook and Twitter, together with the huge amount of data published every day in a variety of different media makes of the Web a highly dynamic repository of any kind of contents. Time dependency in IR is very far from being a solved topic, even though several solutions were proposed. Research works study time dependency in IR for time-sensitiveness of queries, classifying queries for time explicit/implicit needs. To address time freshness, state-of-the-art approaches rank results for the chronological timestamps order of documents in the Trec Microblog Track 2011, or include time as apriori information for probabilistic ranking models based on language models, and so on. We start by studying time dependency/behaviour for machine-learning ranking models, by including static and dynamic time features to explore the importance of time in a microblogging context. Furthermore, datasets and evaluation corpora to understand time dynamics in the ranking process seems inadequate. So, we built a more proper dataset and evaluation corpora by collecting data from the Twitter Microblog service, then we defined a relevance criteria that depends from time. Finally, we present our approach to investigate learning to rank models for Twitter, and preliminary results.