Clifford Xu

Data tools for the China A-share market.

Chinese listed companies file everything to CNINFO, the disclosure platform China's securities regulator designates. It's public, it's free, and it's Chinese PDFs with no structured fields. The scrapers I could find return a title and a link to the PDF.

China A-Share Filings to English (CNINFO) turns those filings into English summaries plus typed figures: dividends, related-party transactions, acquisitions, earnings pre-announcements, share pledges.

Not the extraction. Checking it.

Every figure ships with the Chinese sentence it came from, and the code verifies that sentence against the PDF twice: it has to appear verbatim, and the number has to appear inside it.

The second check exists because of one filing. An early version reported a RMB 1.666 billion shareholder loan from a Vanke announcement, citing a sentence that didn't contain that number at all - the model had cut the sentence short before reaching the figure, and my check only looked for the sentence, so it passed. Truncated citations now get repaired in code.

A dumber one took longer. Four citations kept failing as "not in document" when they obviously were there. The model was writing 3． with a fullwidth period where the PDF had 3.. Same sentence, different byte.

35 of 35 figures pass both checks on the current test set. Failures are flagged, not dropped. If a filing doesn't state something, you get null.

No OCR, so scanned filings are skipped rather than guessed at. Long annual reports get truncated to a character budget, so figures in the cut sections aren't seen. Both are on the Actor page, not buried.

Maths student at the University of Auckland. Open to data engineering roles.

Need a filing type or field I don't cover? Message me.