This article provides a new methodology to predict conflict by using newspaper text. Through machine learning, vast quantities of newspaper text are reduced to interpretable topic shares. We use changes in topic shares to predict conflict one and two years before it occurs. In our predictions we distinguish between predicting the likelihood of conflict across countries and the timing of conflict within each country. Most factors identified by the literature, though performing well at predicting the location of conflict, add little to the prediction of timing. We show that news topics indeed can predict the timing of conflict onset. We also use the estimated topic shares to document how reporting changes before conflict breaks out.

Latent Dirichlet Allocation.