Given a company's financial texts T1..Tn over time, how would you tell whether each one adds new information, and how would you predict next quarter's sales from them?

For each company you have a sequence of financial texts T1, T2, ..., Tn collected over time. Using only these texts, (1) how would you decide whether each Ti carries useful information that the earlier texts did not already contain, and (2) how would you use the texts to predict the company's next quarterly sales? Describe the method, the implementation and the compute it needs, compare alternative approaches, and name the main difficulties and bottlenecks.