In the 1990's at Barra, a Berkeley CA based financial data analysis firm, Daniel Lakeland worked in the Special Projects Research group on a variety of model quality, data quality, and model improvement projects. One product we had jurisdiction over was the Short Term Risk Model (STRM). This model was designed to assess the upcoming variability in stock prices over a few week period and fed its results into other models that needed to know something about the upcoming variability in prices.
One goal for improvement of this model was to incorporate information from the live ticker feed announcements of events such as new products, changes in management, lawsuits, or whatever other announcements related to a given company might appear. Using text processing and classification tools we were able to categorize these announcements into a number of typical categories. Comparing the stock prices on the day of an announcement to the prices on the few days surrounding the announcement allowed us to see if there was any "signal" that could be extracted. The model was not interested in predicting the change in price, but rather the changes in price variability or volatility.
We found that there was evidence of leakage of information in the days preceding the announcement, and that very rapidly after the announcement variability fell back to its predicted values. There was not a lot of signal to be incorporated from the announcement itself, but via appropriate analysis of which events caused persistent changes in volatility over the following days, the model could be tuned to some extent.
This example points out that in some cases, our analysis leads us in new directions. Rather than news being a predictor of volatility, clearly volatility preceded news events due to rumor and leakage of information. Future research might have tried to predict whether observed changes in volatility in the short term could be classified as rumor, and hence expected to persist only for a short time, or as some longer term change in the market conditions for the company. Bringing other sources of information into the model might have allowed this classification of "buzz volatility" to be incorporated into the short term risk predictions.