We analyze the full text of two major British financial publications to measure the evolving structure of London financial news across nearly 140 years. The Financial Times (1888–2008, 3.4 million articles) provides daily coverage from the late Victorian era through the Global Financial Crisis. The Investor's Monthly Manual (1869–1929) provides monthly coverage of the earlier period, including the years 1869–1887 and 1912–1923 missing from the FT digitization.

Using Non-negative Matrix Factorization (NMF) on TF-IDF document-term matrices, we extract interpretable topics and quantify the proportion of news attention allocated to each topic at daily and monthly frequency. Our approach is designed for OCR-degraded historical text: we use FastText subword embeddings that naturally handle garbled 19th-century typography, and track semantic change in financial vocabulary across 12 decades using Procrustes-aligned diachronic embeddings.

Key findings include a 120-year shift from terse market records to narrative-driven analysis — attribution and metaphor rise dramatically — a topic taxonomy whose clusters track major economic shocks, and measurable semantic drift in core financial vocabulary like "gilt," "credit," and "risk." All series are available for download.

Authors

William N. Goetzmann — Yale School of Management

Method

Our methodology is motivated by Bybee, Kelly, Manela, and Xiu (2024), "Business News and Business Cycles," Journal of Finance 79, 3105–3147, who extract topic attention from the Wall Street Journal (1984–2017) and link it to macroeconomic outcomes and asset prices. We adapt their framework to the Financial Times over a much longer horizon (1888–2008), with specific differences noted in each section below.

Financial Times Corpus: 3.4 million articles (1888–1911, 1924–2008). Gap 1912–1923 reflects missing digitization.

Investor's Monthly Manual Corpus: Narrative sections from the IMM (1869–1929), providing coverage of the pre-FT era and the WWI gap.

Topic Model: NMF on TF-IDF (20,000 features, 242 stop words). Bybee et al. use LDA on 180 topics from the WSJ; we use NMF on 50 topics, better suited to the sparser, OCR-degraded historical text. Taxonomy constructed via agglomerative clustering on topic time-series correlations.

Semantic Change: Per-decade FastText (100-dim, skip-gram, char 3-6 n-grams), aligned via orthogonal Procrustes rotation.

Topic Explorer

Select topics to plot their monthly attention over time. Each series shows the average topic weight across all articles published that month.

Source:

News Taxonomy

Topics clustered by the co-movement of their monthly attention, after removing secular trends. We compute first differences of each topic's monthly attention series, then measure the pairwise correlation of these changes. This captures which topics move together in response to events — economic shocks, policy changes, market disruptions — rather than reflecting the long-run structural transition from Victorian to modern financial journalism. The hierarchy is constructed via Ward's agglomerative clustering on the correlation-distance matrix.

Correlation Matrix

Pairwise correlations of first-differenced monthly topic attention. Red = topics whose coverage increases and decreases together in response to events. Blue = topics that substitute for each other — when one gets more attention, the other gets less.

Semantic Change

How has the meaning of financial words changed over 120 years? We train separate FastText embedding models for each decade and align them via Procrustes rotation, then measure semantic drift as cosine distance across time.

Largest Semantic Shifts

Words ranked by the cosine distance between their 1890s and 2000s embeddings. Labels show the decade of the single largest shift.

Track a Word

Select a word to see how its nearest neighbors change across decades. When a word's neighbors shift, its meaning has shifted.

Word Frequency

Monthly frequency (per 1,000 words) of the selected word and its nearest neighbors from the first decade. Reveals whether semantic shifts correlate with changes in usage frequency. 12-month smoothed.

Journalism Style

How does the selected source write, as distinct from what it writes about? We measure six dimensions of journalistic practice using curated word lists, computed independently of the topic model.1 These capture the publication's rhetorical posture: retrospective vs. prospective, attributed vs. asserted, institutional vs. personal, hedged vs. declarative.

Select metrics:

Key Findings — Style

  • Forward-looking language rises from 35% to 49% — the FT shifts from reporting the past to predicting the future
  • Attribution triples (0.5 → 1.5/article) — from unattributed claims to sourced journalism
  • Personalization triples (0.5 → 1.3/article) — from institutional to individual-centered narrative
  • Evaluative language follows a U-shape — judgmental in the Victorian era, neutral mid-century, judgmental again post-1990

Method

Each metric is computed from curated dictionaries of 10–20 words drawn from corpus linguistics (Biber 1988), financial text mining (Loughran & McDonald 2011), and journalism studies (Schudson 1978).2 Ratio metrics show the fraction of relevant words belonging to each pole. Count metrics show occurrences per article. All series are 12-month smoothed.

Narrative Elements

Classic elements of storytelling in financial journalism — conflict, causation, moral judgment, metaphor, suspense, and direct quotation.3 All rise steadily over 120 years. The Victorian FT was a record; the modern FT tells stories.

Select elements:

Key Findings — Narrative

  • Conflict nearly triples (0.29 → 0.81/article) — more adversarial framing
  • Metaphor quadruples (0.07 → 0.25) — "crashed," "soared," "bubble," "meltdown"
  • Moral/blame rises (0.38 → 0.54) — more judgment of responsibility
  • Suspense rises (0.24 → 0.39) — "remains unclear," "crucial question"

Key Findings — Historical Memory

  • Historical references rise from 0.8 to 2.0 per article — journalism becomes more historically referential
  • The 1990s obsessively reference 1987 (Black Monday); the 1930s reference 1929
  • "Great Depression" as a term barely exists until the 1980s — it is a retrospective label, not a contemporary one

Historical Memory

How often does the FT look backward? We count references to specific years more than two years in the past. Financial journalism has become increasingly historically referential — the modern FT frames current events through comparison with prior crises, booms, and policy episodes.

Notes

1 Style dimensions. Forward-looking vs. retrospective: Barnhurst & Mutz (1997), J. Communication 47(4), 27–53; Fink & Schudson (2014), Journalism 15(1), 3–20. Attribution: Tuchman (1978), Making News (Free Press); Schudson (2001), Journalism 2(2), 149–170. Personalization: Van Aelst, Sheafer & Stanyer (2012), Journalism 13(2), 203–220. Hedged vs. declarative: Salgado & Stromback (2012), Journalism 13(2), 144–161. Evaluative language: Wahl-Jorgensen (2013), Journalism 14(1), 129–145; Entman (1993), J. Communication 43(4), 51–58. Financial journalism history: Barnhurst & Nerone (2001), The Form of News (Guilford); Schifferes & Roberts, eds. (2015), The Media and Financial Crises (Routledge); Starkman (2014), The Watchdog That Didn't Bark (Columbia). ↩

2 Dictionaries and measurement. Biber (1988), Variation across Speech and Writing (Cambridge); Loughran & McDonald (2011), J. Finance 66(1), 35–65; Schudson (1978), Discovering the News (Basic Books); Bednarek & Caple (2017), The Discourse of News Values (Oxford); Gentzkow, Kelly & Taddy (2019), J. Economic Literature 57(3), 535–574. ↩

3 Narrative categories. General: Bruner (1991), Critical Inquiry 18(1), 1–21; Schudson (1982), Daedalus 111(4), 97–112; Shiller (2017), American Economic Review 107(4), 967–1004; Roos & Reccius (2024), J. Economic Surveys 38(2), 303–341. Conflict and causation: Entman (1993), J. Communication 43(4), 51–58; Semetko & Valkenburg (2000), J. Communication 50(2), 93–109; Valkenburg, Semetko & de Vreese (1999), Communication Research 26(5), 550–569. Metaphor: Lakoff & Johnson (1980), Metaphors We Live By (Chicago); McCloskey (1985), The Rhetoric of Economics (Wisconsin); Arrese & Vara-Miguel (2016), Discourse & Society 27(2), 133–155. Moral/blame: Iyengar (1991), Is Anyone Responsible? (Chicago). Sentiment: Tetlock (2007), J. Finance 62(3), 1139–1168; Garcia (2013), J. Finance 68(3), 1267–1300. ↩

Downloadable Data

FileDescriptionDownload
Monthly Topic Attention Monthly average topic weights (theta matrix), 1888–2008 CSV
Daily Topic Attention Daily average topic weights CSV
Topic Words Top 20 words per topic with weights CSV
Topic Metadata Labels, cluster assignments, top words (JSON) JSON
Semantic Change Diachronic shift data and neighbors by decade JSON