In this paper, we study news group modeling and forecasting methods using quantitative data generated by our large-scale natural language processing (NLP) text analysis system. A news group is a set of news entities, like top U.S. cities, governors, senators, golfers, or movie actors. Our fame distribution analysis of news groups shows that log-normal and power-law distributions generally could describe news groups in many aspects. We use several real news groups including cities, politicians, and CS professors, to evaluate our news group models in terms of time series data distribution analysis, group-fame probability analysis, and fame-changing analysis over long time. We also build a practical news generation model using a HMM (Hidden Markov Model) based approach. Most importantly, our analysis shows the future entity fame distribution has a power-law tail. That is, only a small number of news entities in a group could become famous in the future. Based on these analysis we are able to answer some interesting forecasting problems - for example, what is the future average fame (or maximum fame) of a specific news group? And what is the probability that some news entity become very famous within a certain future time range? We also give concrete examples to illustrate our forecasting approaches.