Forecasting societal events such as civil unrest, mass protests, and violent conflicts is a challenging problem with several important real-world applications in planning and policy making. While traditional forecasting approaches have typically relied on historical time series for generating such forecasts, recent research has focused on using open source surrogate data for more accurate and timely forecasts. Furthermore, leveraging such data can also help to identify precursors of those events that can be used to gain insights into the generated forecasts. The key challenge is to develop a unified framework for forecasting and precursor identification that can deal with missing historical data. Other challenges include sucient flexibility in handling different types of events and providing interpretable representations of identified precursors. Although existing methods exhibit promising performance for predictive modeling in event detection, these models do not adequately address the above challenges. Here, we propose a unified framework based on an attention-based long short-term memory (LSTM) model to simultaneously forecast events with sequential text datasets as well as identify precursors at different granularity such as documents and document excerpts. The key idea is to leverage word context in sequential and time-stamped documents such as news articles and blogs for learning a rich set of precursors. We validate the proposed framework by conducting extensive experiments with two real-world datasets—military action and violent conflicts in the Middle East and mass protests in Latin America. Our results show that overall, the proposed approach generates more accurate forecasts compared to the existing state-of-the-art methods, while at the same time producing a rich set of precursors for the forecasted events.