This document discusses using news summaries to model socioeconomic patterns and jointly model the importance of text and news outlets in forecasts. It describes extracting daily summaries about the EU from the OpenEuropeThinkTank in English, averaging 14 news per summary from 435 outlets from 2006 to 2013. Models using linear and bilinear regressions that jointly consider word frequencies and outlet weights were able to predict economic sentiment indicator values with errors of 9.89% and 8.77%, slightly better than linear regression alone.