FU902: Seminar 04
David Procházka
2023-11-19
1 Recap the basics of the regression modelling in R
2 Read relevant literature
Loughran, T., & McDonald, B. (2011). When is a liability not a liability? Textual analysis, dictionaries, and 10‐Ks. The Journal of finance, 66(1), 35-65. https://doi.org/10.1111/j.1540-6261.2010.01625.x
Loughran, T., & McDonald, B. (2015). The use of word lists in textual analysis. Journal of Behavioral Finance, 16(1), 1-11. https://doi.org/10.1080/15427560.2015.1000335
Lang, M., & Stice-Lawrence, L. (2015). Textual analysis and international financial reporting: Large sample evidence. Journal of Accounting and Economics, 60(2-3), 110-135. https://doi.org/10.1016/j.jacceco.2015.09.002
Guo, L., Shi, F., & Tu, J. (2016). Textual analysis and machine leaning: Crack unstructured data in finance and accounting. The Journal of Finance and Data Science, 2(3), 153-170. https://doi.org/10.1016/j.jfds.2017.02.001
Loughran, T., & McDonald, B. (2016). Textual analysis in accounting and finance: A survey. Journal of Accounting Research, 54(4), 1187-1230. https://doi.org/10.1111/1475-679x.12123
Bochkay, K., Brown, S. V., Leone, A. J., & Tucker, J. W. (2023). Textual analysis in accounting: What’s next?. Contemporary accounting research, 40(2), 765-805. https://doi.org/10.1111/1911-3846.12825
3 Learn (textual analysis) modelling by example
A. Download and import the datasheet
All files are available at this shared folder
For this Tutorial, the sub-folder S04 is relevant; download the txt files to your local computer. Please note that the files are the Management Discussion and Analysis by US listed firms. The files were extracted from the EDGAR database using the R-package EDGAR
Set a working directory using the command setwd(“C:/…/…/…/…”); where you have the files downloaded (stored) and where outputs will be exported (e.g., setwd(“C:/Users/MyPC/Documents/R/Paper01”)
Read the txt files from the working directory
library(quanteda)
docs <- list.files(path="C:/Users/MyPC/Documents/R/Paper01", pattern=".txt", all.files=TRUE, full.names=TRUE)- Convert the files into a corpus
library(readtext)
corpus2019 <- corpus(readtext(docs))B. Analyse the data
- Inspect the corpus
summary(corpus2019)## Corpus consisting of 30 documents, showing 30 documents:
##
## Text Types Tokens Sentences
## ABB.txt 2191 11734 336
## ALPHABET.txt 1730 10806 304
## AMAZON.txt 1547 9359 267
## ANTHEM.txt 2673 20146 591
## APPLE.txt 1273 5886 155
## BOEING.txt 2350 15144 497
## CATERPILLAR.txt 3338 26786 850
## CISCO.txt 2201 14270 387
## COCACOLA.txt 3178 31491 986
## COMCAST.txt 2318 16984 385
## CONOCO.txt 3338 26223 879
## COSTCO.txt 1291 5975 193
## DISNEY.txt 2316 18861 418
## EXXON.txt 3185 16939 576
## FACEBOOK.txt 1745 11375 325
## HOMEDEPOT.txt 1182 5877 206
## IBM.txt 896 5902 160
## JOHNSON.txt 2398 13456 427
## LOCKHEED.txt 2895 26751 843
## LOWES.txt 1954 12748 378
## MERCK.txt 3142 26948 834
## NIKE.txt 2125 14331 433
## ORACLE.txt 2086 15634 347
## PEPSICO.txt 4195 44393 1111
## PROCTER.txt 2482 18564 652
## TESLA.txt 4473 48860 1368
## TMOBILE.txt 4008 36932 919
## UNITEDHEALTH.txt 1788 7936 228
## UPS.txt 2831 24017 704
## WALMART.txt 1837 11890 322Note: If you like to see the text of the report, e.g. for AMAZON, use the code as.character(corpus2019)[“AMAZON.txt”] or as.character(corpus2019)[3].
- Reorder the corpus based on the number of words (tokens)
tokeninfo2019 <- summary(corpus2019)
tokeninfo2019$Text <- with(tokeninfo2019, reorder(Text, -Tokens))- Plot the length of the documents (based on the number of tokens)
library(ggplot2)
ggplot(tokeninfo2019, aes(x = Text, y = Tokens)) +
geom_point() +
theme(axis.text.x = element_text(angle = 90, hjust = 1)) + labs(x = "FIRM", y = "MD&A words length") + geom_hline(yintercept = mean(tokeninfo2019$Tokens), colour = "#b10061")
- Find all occurrences of the word “merger”, including its context
tokens2019 <- tokens(corpus2019)
kwic(tokens2019, pattern = "*merger*")## Keyword-in-context with 57 matches.
## [ANTHEM.txt, 3312] the Agreement and Plan of | Merger
## [ANTHEM.txt, 3316] of Merger, or Cigna | Merger
## [ANTHEM.txt, 3357] the terms of the Cigna | Merger
## [ANTHEM.txt, 3371] litigation related to the Cigna | Merger
## [ANTHEM.txt, 3389] Regulatory Proceedings - Cigna Corporation | Merger
## [ANTHEM.txt, 4385] as a result of strategic | mergers
## [ANTHEM.txt, 8876] related to the terminated Cigna | Merger
## [ANTHEM.txt, 9060] related to the terminated Cigna | Merger
## [ANTHEM.txt, 9298] related to the terminated Cigna | Merger
## [COCACOLA.txt, 11692] , capital improvement programs, | merger
## [IBM.txt, 1432] . Agreement and Plan of | Merger
## [PEPSICO.txt, 2923] otherwise noted, reflect all | mergers
## [PEPSICO.txt, 3929] impairment charges 0.18 0.16 0.09 | Merger
## [PEPSICO.txt, 5028] gains or losses associated with | mergers
## [PEPSICO.txt, 5387] and 2014 Productivity Plans, | merger
## [PEPSICO.txt, 6384] 272 36 56 1 251 | Merger
## [PEPSICO.txt, 7458] our consolidated financial statements. | Merger
## [PEPSICO.txt, 7467] In 2018, we incurred | merger
## [PEPSICO.txt, 10764] c ) 63 53 60 | Merger
## [PEPSICO.txt, 11056] Table of Contents and the | merger
## [PEPSICO.txt, 14681] impairment charges 0.4 0.3 0.1 | Merger
## [PEPSICO.txt, 39852] Stock In connection with our | merger
## [PEPSICO.txt, 41159] accounting for business combinations, | merger
## [PEPSICO.txt, 41192] In 2018, we incurred | merger
## [PEPSICO.txt, 43641] Acquisitions and divestitures: all | mergers
## [TMOBILE.txt, 136] as a result of the | Merger
## [TMOBILE.txt, 168] and broadband industries. The | Merger
## [TMOBILE.txt, 4007] expenses. Joint ventures, | mergers
## [TMOBILE.txt, 7347] of spectrum; engaging in | mergers
## [TMOBILE.txt, 8504] acquisition, investment, or | merger
## [TMOBILE.txt, 8529] of, investments in or | mergers
## [TMOBILE.txt, 8836] acquisition, investment, or | merger
## [TMOBILE.txt, 12022] incorporation, a sale or | merger
## [TMOBILE.txt, 12279] conclusion of the proposed Sprint | Merger
## [TMOBILE.txt, 13014] consolidation, or other strategic | mergers
## [TMOBILE.txt, 13043] activities, including the pending | Merger
## [TMOBILE.txt, 13056] likelihood the conditions to the | Merger
## [TMOBILE.txt, 13066] the anticipated benefits of the | Merger
## [TMOBILE.txt, 13381] stock repurchase program if the | Merger
## [TMOBILE.txt, 14110] after giving effect to the | Merger
## [TMOBILE.txt, 14306] . Failure to complete the | Merger
## [TMOBILE.txt, 14332] of operations. If the | Merger
## [TMOBILE.txt, 14373] a market assumption that the | Merger
## [TMOBILE.txt, 14455] is expected that if the | Merger
## [TMOBILE.txt, 14497] . Further, if the | Merger
## [TMOBILE.txt, 16064] the full benefits of the | Merger
## [TMOBILE.txt, 16147] projected accretive effect of the | Merger
## [TMOBILE.txt, 16160] our common stock following the | Merger
## [TMOBILE.txt, 16275] expect to conduct, certain | pre-Merger
## [TMOBILE.txt, 16306] After giving effect to the | pre-Merger
## [TMOBILE.txt, 17120] of spectrum; engaging in | mergers
## [TMOBILE.txt, 17454] long-term financings prior to the | merger
## [TMOBILE.txt, 19987] as a result of the | Merger
## [TMOBILE.txt, 20021] industries. Immediately following the | Merger
## [TMOBILE.txt, 20096] 31, 2018. The | Merger
## [TMOBILE.txt, 31135] to the consummation of the | Merger
## [TMOBILE.txt, 31188] Comprehensive Income. If the | Merger
##
## | , or Cigna Merger Agreement
## | Agreement, between us and
## | Agreement. For additional information
## | Agreement, see Note 13
## | Litigation, of the Notes
## | and acquisitions, as well
## | Agreement. Our selling,
## | Agreement. For the year
## | Agreement, the excess tax
## | and acquisition plans, and
## | , dated as of October
## | and acquisitions activity, including
## | and integration charges 0.05 Net
## | , acquisitions, divestitures and
## | and integration charges associated with
## | and integration charges ( 75
## | and Integration Charges In 2018
## | and integration charges of $
## | and integration charges ( c
## | and integration charges related to
## | and integration charges 0.1 Net
## | with The Quaker Oats Company
## | and integration costs are not
## | and integration charges of $
## | and acquisitions activity, including
## | , is expected to be
## | is subject to regulatory approvals
## | , acquisitions and strategic alliances
## | , acquisitions, business combinations
## | may subject us to significant
## | with businesses, technologies,
## | may have a material adverse
## | of our Company and other
## | . The current royalty rate
## | or acquisition activities involving us
## | and views of market participants
## | will be satisfied and the
## | will be realized; disruptions
## | fails to close. Our
## | ) from at least two
## | could negatively impact us and
## | is not completed for any
## | will be completed. In
## | is not completed, we
## | is not completed, it
## | may not be realized,
## | , and negatively impact the
## | . As a result,
## | financing transactions, which will
## | financing transactions and the Transactions
## | , acquisitions, business combinations
## | closing. However, there
## | , is expected to be
## | , it is anticipated that
## | is subject to regulatory approvals
## | . 47 Table of Contents
## | is consummated, we will- Find 50 most frequent words
dfm2019 <- corpus2019 |> tokens(remove_punct = TRUE, remove_symbols = TRUE, remove_numbers = TRUE) |>
tokens_remove(stopwords("en")) |> dfm()
topfeatures(dfm2019, 50)## net million tax sales operating financial
## 2329 2290 2036 2015 1905 1883
## cash due billion costs income primarily
## 1881 1735 1705 1650 1599 1475
## increase related increased revenue december company
## 1345 1313 1277 1273 1269 1256
## business may consolidated year operations fiscal
## 1216 1178 1158 1138 1097 1086
## assets certain rate total statements value
## 1079 1052 1052 1050 1008 986
## including higher expense impact results cost
## 984 977 969 968 958 949
## u.s products compared interest expenses revenues
## 941 935 932 898 886 882
## growth changes debt offset future table
## 879 832 826 802 779 749
## services activities
## 743 723- Plot 50 most frequent words
library(quanteda.textstats)
features2019 <- textstat_frequency(dfm2019, n = 50)
features2019$feature <- with(features2019, reorder(feature, -frequency))
ggplot(features2019, aes(x = feature, y = frequency)) +
geom_point(colour = "#b10061") + labs(x = "Word", y = "Frequency") +
theme(axis.text.x = element_text(angle = 90, hjust = 1))
- Find 50 most frequent bigrams
library(tidytext)
library(dplyr)
Bigram2019 <- tidy(dfm2019) %>% unnest_tokens(output = bigram, input = term, token = "ngrams", n = 2)
print(Bigram2019 %>% count(bigram, sort=TRUE), n = 50)## # A tibble: 1,331 × 2
## bigram n
## <chr> <int>
## 1 <NA> 45404
## 2 10 k 30
## 3 long term 30
## 4 short term 24
## 5 off balance 23
## 6 third party 23
## 7 2019 02 18
## 8 non gaap 18
## 9 one time 18
## 10 non cash 16
## 11 indefinite lived 15
## 12 weighted average 15
## 13 forward looking 14
## 14 over year 14
## 15 year end 14
## 16 year over 14
## 17 long lived 13
## 18 stock based 13
## 19 multi year 12
## 20 percentage point 12
## 21 pre tax 12
## 22 share based 12
## 23 straight line 12
## 24 five year 11
## 25 fixed rate 10
## 26 for sale 10
## 27 high quality 10
## 28 near term 10
## 29 non u.s 10
## 30 three year 10
## 31 available for 9
## 32 write down 9
## 33 364 day 8
## 34 after tax 8
## 35 to market 8
## 36 write downs 8
## 37 low taxed 7
## 38 non recurring 7
## 39 self insurance 7
## 40 year to 7
## 41 1 3 6
## 42 2014 09 6
## 43 8 k 6
## 44 full year 6
## 45 mark to 6
## 46 non operating 6
## 47 one year 6
## 48 self insured 6
## 49 4 5 5
## 50 a 1 5
## # ℹ 1,281 more rows- Find 25 most frequent trigrams
Trigram2019 <- tidy(dfm2019) %>% unnest_tokens(output = bigram, input = term, token = "ngrams", n = 3)
print(Trigram2019 %>% count(bigram, sort=TRUE), n = 25)## # A tibble: 185 × 2
## bigram n
## <chr> <int>
## 1 <NA> 47210
## 2 year over year 14
## 3 available for sale 9
## 4 mark to market 6
## 5 period to period 5
## 6 year to year 5
## 7 direct to consumer 4
## 8 other than temporary 4
## 9 2019 02 27 3
## 10 build to suit 3
## 11 end to end 3
## 12 over the counter 3
## 13 2019 01 31 2
## 14 2019 02 08 2
## 15 2019 02 19 2
## 16 2019 02 20 2
## 17 2019 02 21 2
## 18 2019 03 28 2
## 19 go to market 2
## 20 hart scott rodino 2
## 21 out of date 2
## 22 period over period 2
## 23 point of sale 2
## 24 right of use 2
## 25 year on year 2
## # ℹ 160 more rows- Plot the dispersion of the terms “cash flow”, “income”, “profit” within the documents
library(quanteda.textplots)
idx1 <- index(tokens(corpus2019), phrase ("cash flow"))
idx2 <- index(tokens(corpus2019), phrase ("income"))
idx3 <- index(tokens(corpus2019), phrase ("profit"))
textplot_xray(kwic(tokens(corpus2019), index = idx1), kwic(tokens(corpus2019), index = idx2), kwic(tokens(corpus2019), index = idx3), sort = TRUE)
- Identify 10 most frequently occurring topics
library(stm)
quant2019 <- dfm_trim(tokens(corpus2019, remove_punct = TRUE, remove_numbers = TRUE) |>
tokens_remove(stopwords("en")) |>
dfm(), min_termfreq = 10)
set.seed(100)
plot(stm(quant2019, K = 10, verbose = FALSE))
- Visualise the connections among words
network2019 <-
tokens(corpus2019, remove_punct = TRUE) %>%
tokens_tolower() %>%
tokens_remove(stopwords("english"), padding = FALSE) %>%
fcm(context = "window", window = 5, tri = FALSE)
topfeats2019 <- names(topfeatures(network2019, 50))
set.seed(100)
textplot_network(fcm_select(network2019, topfeats2019), min_freq = 10, edge_color = "#b10061") + ggtitle("Words network") 
- Calculate (cosine) similarity between each pair of the documents
dfmat2019 <- corpus2019 |>
tokens(remove_punct = TRUE, remove_symbols = TRUE, remove_numbers = TRUE) |>
tokens_wordstem(language = "en") |>
tokens_remove(stopwords("en")) |>
dfm()
tstat2019 <- textstat_simil(dfmat2019, margin = "documents", method = "cosine")- Extract the similarity indices for the APPLE company and visualise them
as.list(tstat2019)$"APPLE"## JOHNSON.txt MERCK.txt LOWES.txt PROCTER.txt
## 0.7610193 0.7014893 0.6778258 0.6666915
## AMAZON.txt COCACOLA.txt COSTCO.txt PEPSICO.txt
## 0.6639099 0.6537156 0.6523390 0.6503047
## ABB.txt CATERPILLAR.txt WALMART.txt CISCO.txt
## 0.6407577 0.6386842 0.6358225 0.6117830
## FACEBOOK.txt LOCKHEED.txt HOMEDEPOT.txt ORACLE.txt
## 0.5918962 0.5886400 0.5776497 0.5740082
## TMOBILE.txt CONOCO.txt ANTHEM.txt NIKE.txt
## 0.5687208 0.5652326 0.5606723 0.5432352
## DISNEY.txt COMCAST.txt BOEING.txt UNITEDHEALTH.txt
## 0.5367939 0.5097931 0.5087365 0.5054977
## UPS.txt ALPHABET.txt TESLA.txt EXXON.txt
## 0.4889719 0.4720312 0.4612674 0.4224644
## IBM.txt
## 0.2208358dotchart(as.list(tstat2019)$"APPLE", cex = 0.6, pt.cex = 1.0, pch = 19, lcolor = "#b10061")
- Convert Document-feature matrix (quanteda package) to Document-term matrix (tm package)
library(tm)
dtm2019 <- as.DocumentTermMatrix(dfm2019, weighting = weightTf)- Tokenise the document-term matrix
library(tidytext)
td2019 <- tidy(dtm2019)- Assess the sentiment of the reports, using a general lexicon, e.g. Bing
library(dplyr)
sentiments_Bing <- td2019 %>%
inner_join(get_sentiments("bing"), by = c(term = "word"))
sentiments_Bing## # A tibble: 3,842 × 4
## document term count sentiment
## <chr> <chr> <dbl> <chr>
## 1 ABB.txt dedicated 1 positive
## 2 COCACOLA.txt dedicated 1 positive
## 3 EXXON.txt dedicated 1 positive
## 4 PROCTER.txt dedicated 1 positive
## 5 TMOBILE.txt dedicated 1 positive
## 6 UNITEDHEALTH.txt dedicated 2 positive
## 7 ABB.txt innovation 2 positive
## 8 CISCO.txt innovation 2 positive
## 9 COCACOLA.txt innovation 5 positive
## 10 JOHNSON.txt innovation 1 positive
## # ℹ 3,832 more rows- Visualise the most frequent positive and negative words for the Bing lexicon
sentiments_Bing %>%
count(sentiment, term, wt = count) %>%
ungroup() %>%
filter(n >= 100) %>%
mutate(n = ifelse(sentiment == "negative", -n, n)) %>%
mutate(term = reorder(term, n)) %>%
ggplot(aes(term, n, fill = sentiment)) +
geom_bar(stat = "identity") +
scale_fill_manual(values=c("black",
"#b10061")) +
xlab("") +
ylab("Contribution to sentiment") +
coord_flip()
Note: Are the terms such as “cloud” or “gross” really negative in the context of financial reporting?
- Assess the sentiment of the reports, using a financial lexicon, e.g., that developed by Loughran&McDonald (2011)
library(dplyr)
sentiments_Loughran <- td2019 %>%
inner_join(get_sentiments("loughran"), by = c(term = "word"))
sentiments_Loughran## # A tibble: 5,264 × 4
## document term count sentiment
## <chr> <chr> <dbl> <chr>
## 1 ABB.txt accession 1 litigious
## 2 ALPHABET.txt accession 1 litigious
## 3 AMAZON.txt accession 1 litigious
## 4 ANTHEM.txt accession 1 litigious
## 5 APPLE.txt accession 1 litigious
## 6 BOEING.txt accession 1 litigious
## 7 CATERPILLAR.txt accession 1 litigious
## 8 CISCO.txt accession 1 litigious
## 9 COCACOLA.txt accession 1 litigious
## 10 COMCAST.txt accession 1 litigious
## # ℹ 5,254 more rows- Visualise the most frequent positive and negative words for the Loughran&McDonald lexicon
sentiments_Loughran %>%
count(sentiment, term, wt = count) %>%
ungroup() %>%
filter(n >= 100) %>%
mutate(n = ifelse(sentiment == "negative", -n, n)) %>%
mutate(term = reorder(term, n)) %>%
ggplot(aes(term, n, fill = sentiment)) +
geom_bar(stat = "identity") +
xlab("") +
ylab("Contribution to sentiment") +
coord_flip()
- or alternatively
sentiments_Loughran %>%
count(term, sentiment, sort = TRUE) %>%
group_by(sentiment) %>%
top_n(10) %>%
ungroup() %>%
mutate(term = reorder(term, n)) %>%
ggplot(aes(term, n, fill = sentiment)) +
geom_col(show.legend = FALSE) +
facet_wrap(~sentiment, scales = "free_y") +
labs(y = "Contribution to sentiment",
x = NULL) +
coord_flip()
- Extract sentiment data by firm for a follow-up analysis
group_by(sentiments_Bing %>% count(sentiment, document, wt = count), document, sentiment) %>% summarise(words = sum(n)) # Bing## # A tibble: 60 × 3
## # Groups: document [30]
## document sentiment words
## <chr> <chr> <dbl>
## 1 ABB.txt negative 117
## 2 ABB.txt positive 250
## 3 ALPHABET.txt negative 70
## 4 ALPHABET.txt positive 113
## 5 AMAZON.txt negative 84
## 6 AMAZON.txt positive 199
## 7 ANTHEM.txt negative 219
## 8 ANTHEM.txt positive 437
## 9 APPLE.txt negative 87
## 10 APPLE.txt positive 55
## # ℹ 50 more rowsgroup_by(sentiments_Loughran %>% count(sentiment, document, wt = count), document, sentiment) %>% summarise(words = sum(n)) # Loughran## # A tibble: 165 × 3
## # Groups: document [30]
## document sentiment words
## <chr> <chr> <dbl>
## 1 ABB.txt constraining 68
## 2 ABB.txt litigious 43
## 3 ABB.txt negative 100
## 4 ABB.txt positive 119
## 5 ABB.txt superfluous 8
## 6 ABB.txt uncertainty 154
## 7 ALPHABET.txt constraining 32
## 8 ALPHABET.txt litigious 22
## 9 ALPHABET.txt negative 80
## 10 ALPHABET.txt positive 49
## # ℹ 155 more rows