The first event study was very simple and broad, designed to compare with Tetlock,
Saar-Tsechansky, and Macskassy’s earlier result (seen in Figure 6.1). It is shown in
Figures 6.8a, b and is indeed very similar. In sentiment metrics work the positive event
lines are consistently above the negative event lines. Timeliness is an issue (as seen in
Figure 6.8a), with large pre-event returns observed, but potentially exploitable post-
event returns are also seen (as in Figure 6.8b).
6.4.3 Adjusting aggregate event parameters and thresholds, and segmentation by sector
These give promising results for the full S&P 1500 index (see Figure 6.9), but need
refinement to produce alpha in excess of reasonable transaction costs. Our approach was
to segment by sector, and adjust the news analytic filter settings.
One observation in all of these event studies was that the positive sentiment return
lines are consistently above the negative sentiment returns, and opposite in direction.
The sentiment measures appear very effective. In addition, changes in traffic/sentiment
thresholds show expected effects. More stringent filters reduce the numbers of events,
but are associated with larger excess returns.
The best sectors for this approach are: basic materials, cyclicals, financials, indus-
trials, non-cyclicals, and technology. These effects are illustrated for financials and non-
cyclicals in Figures 6.10 and 6.11.
6.4.4 Adjusting sentiment thresholds
The effect seen from 2003 through 2008 is evident in 2009 (through the end of Q3), with
about the same magnitude. We still observe the balance between breadth of signal and
magnitude of excess return, although at very extreme levels there are two few data points
to draw any conclusions.
Relating news analytics to stock returns 157Figure 6.7.Monthly difference in number of positive and negative news items.