wsjibm: WSJ Stories on IBM

Description

Word counts for Wall Street Journal story abstracts with IBM in the title, along with the concurrent returns on IBM stock.

Arguments

Value

wsjibmCountsA simple_triplet_matrix of counts indexed by article-rows and word-columns.
wsjibmReturnsA matrix containing the corresponding publication DATE along with IBM's two-day holding returns (RET) and return over the S&P500 (ROM).

Details

Headlines and one-sentence abstracts for Wall Street Journal (WSJ) stories with IBM in the headline, dating from August 1988 to August 2010, were retrieved from the ProQuest database. Each article is accompanied by two-day return and return-over-market for shares in IBM listed on the New York Stock Exchange, calculated from the opening of the previous day to market close on the day of publication. Full details are available in Taddy (2011).

References

Taddy (2011), Inverse Regression for Analysis of Sentiment in Text. http://arxiv.org/abs/1012.2098

Examples

Run this code

data(wsjibm)
## fit a simple finite mixture model (not the usual mixed-membership admix)
## allow different topic-membership probabilities by gain/loss
newstpx <- topics(wsjibmCounts, K=5, admix=FALSE, grp=wsjibmReturns$ROM>=0) 
plot(newstpx,3, col=3, cex.lgdc=.6, ylab="gain")
summary(newstpx, nwrd=10)

## fit admixture topics over years, using prior shape to allow them to change in time
year <- factor(1900 + as.POSIXlt(wsjibmReturns$DATE)$year)
Y <- nlevels(year)
annualtopics <- vector(length=Y, mode="list")
topwords <- c()
shape=NULL
for(i in 1:Y){ 
      annualtopics[[i]] <- topics(wsjibmCounts[year==levels(year)[i],], K=5, shape=shape)
      topwords <- cbind(topwords, as.character(summary(annualtopics[[i]], verb=FALSE)$phrase))
      delta <- 10000 # weight of the previous year in number of words observed per topic		
      shape <- annualtopics[[i]]$theta*delta }
## top 5 words by topic in past 4 years
dimnames(topwords) <- list(topic=rep(1:5,each=5), year=levels(year))
print(topwords[,Y - 3:0])