Private preview — please enter the access password.
This project applies the computational-history toolkit of the Structure of London Financial News to the Shanghai press. The first corpus is the North China Herald (1850–1941), the leading English-language weekly of treaty-port China. The full run of the paper was OCR'd, segmented into articles, and modelled with Non-negative Matrix Factorization (NMF) on a TF-IDF document–term matrix.
Unlike the London project, we model the whole paper and then carry a finance flag — a commercial/financial lexicon scored per article — up to the topic level, so each topic reports a finance share. This lets the financial structure of the news surface from a general newspaper, and sets up direct comparison with the Chinese-language Shenbao 申報, to be added next.
Loading corpus statistics…
Motivated by Bybee, Kelly, Manela & Xiu (2024), "Business News and Business Cycles," Journal of Finance, who extract topic attention from the press and link it to the economy.
Corpus: North China Herald, Internet Archive collection
northchinaherald (public domain); IA Tesseract OCR.
Pipeline: issue download → article segmentation by section headers → finance lexicon flag → NMF (50 topics) on TF-IDF, with a dictionary/domain OCR-junk vocabulary filter → finance share per topic + decade timeline.
Caveat: historical newsprint OCR is imperfect; the finance flag is a transparent keyword heuristic, not a trained classifier.
Yearly attention to each NMF topic across the paper's run. Lines break at gap years (1851–55, 1894–95) where issues are missing from the archive. Topics are tagged with their finance share.
How topics relate. The dendrogram clusters topics by the correlation of their yearly attention; the heatmap shows the full topic–topic correlation matrix. Financial topics are shaded red in the tree.
How financial was the paper over time, and which financial themes dominated. Two cross-validating signals: the share of articles flagged finance by a commercial lexicon, and the finance-weighted topic attention.
All series are available as JSON. The corpus is built from the public-domain Internet Archive North China Herald collection.