To establish the empirical significance of our news indices, we use the event study methodology (see Campbell, Lo, and MacKinlay, 1997 for background on event studies). We review the basics of this well-known technique in Section 3.5.1, and provide a few illustrative examples in Section 3.5.2. We present formal statistical tests for the significance of news index events in Sections

3.5.1 Event analysis

For a given index, event analysis is performed in the following manner. We compute the index over the sample period from January 1, 2003 to March 31, 2007, and declare that an “event” has taken place whenever the score exceeds a certain threshold, typically 0:995. We then remove any event that follows less than 1 hour after another event, which guards against having many events in quick succession that all reflect the same news event. We then analyze the behavior of exchange rates in the periods before and after these events.

In our analysis, we focus on two time-series describing the behavior of exchange rates.

The first is the time-series of log returns, denoted {ri}i. Since we only have banks' quote data, this series is derived by taking the logarithm of the geometric mean of bid and ask quotes (as described in Section 3.3). The second time-series we consider is that of de-seasonalized squared log returns, denoted {si}i, which is a measure of volatility in exchange rates. Since exchange rate volatilities exhibit strong weekly seasonalities ...

Get The Handbook of News Analytics in Finance now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.