We analyze the network structure of lagged correlations among daily financial news sentiments and returns of financial market indices of 40 countries from 2002 to 2012. Using a spectral method, we decompose the network into bipartite sub-structures, and show that these sub-structures are relevant to the performance of prediction models, bridging concepts from network theory and time series analysis. Our results suggest that, at the daily level, endogenous influences among financial markets overwhelm exogenous influences of news outlets, and that changes in financial news sentiments respond to market movements more substantially than they drive them.