Country Risk

How closely it matches the published series

Computed against the CountryRisk series in the published firm-country-quarter panel distributed at country-risk.net on the matched overlap ($N = 321,048$ matched firm-quarters per country across all 76 quarters from 2002-Q1 through 2020-Q4):

At the firm level, 43 of the 45 countries exceed a correlation of 0.85; the two exceptions—Israel (0.748) and Taiwan (0.786)—correlate at 0.932 and 0.965 after quarterly aggregation. The comparison scales call-level scores by transcript word count, collapses to quarterly firm means, and carries gaps forward as in the original import pipeline. The MIT-licensed script in our countryrisk-correlations GitHub repository reproduces every number in this table from the firm-level files on this page and the public archive at country-risk.net.

What is the same, and what is different

Verified directly against the original implementation, the following are identical:

• Scoring rule: Every occurrence of a country term is a center, risk or uncertainty synonyms are counted in the ±10-word window around it (with the country term's own tokens included), and each occurrence contributes the country term's tf-idf weight multiplied by that risk-word count.
• Lexicons: The per-country tf-idf keyword libraries, weights, and Oxford Thesaurus risk/uncertainty synonym set match the published paper.
• Masking and case rules: Window geometry for one- and two-word country names, case-sensitive country matching, per-occurrence accumulation, and special-case masking (excluding "Gulf of Mexico" from Mexico) match the original code.

Remaining differences affect sample mapping and downstream scaling rather than the scoring formula:

• Tokenization: Punctuation-split word boundaries with digits retained; a small number of de-accented institutional terms and rare 3+ word entries do not match, with negligible effect on earnings calls.
• Sample composition: Updated mapping to gvkey and headquarters country accounts for most of the remaining call-level difference and averages out in quarterly aggregates.
• Transcript length and safe-harbor rules: No minimum-length filter on very short transcripts, and minor differences in one safe-harbor boilerplate rule.
• Scale and panel structure: Each CSV row is one earnings call in one country series with earningscallID, calendar date, UTC start time, company identifiers (gvkey, isin, tickersymbol, headquarters country), raw weighted risk score, and transcript word count (nr_of_words). To match the published panel's scale, multiply the raw score by 100,000, divide by nr_of_words, average to firm-quarter, and apply your preferred gap-fill rule.