Decoding RaymondsIndex: Why "Accurate" and "Useful" Are Different Tests

On 16 June 2026, Korea's Financial Supervisory Service released its analysis of FY2025 audit opinions for listed companies. Most coverage led with the alarming number: among companies that received a clean opinion but carried a going-concern emphasis paragraph, 32.1% were delisted or received a modified opinion the following year.

It is an alarming number. It is also the least interesting thing in the release.

The opinion line is not the signal

Of 2,702 listed companies analysed, 2,637 — 97.6% — received an unqualified opinion. That figure has not left a narrow band since Korea's 2019 external audit reform: 97.2% in 2021, 97.9% in 2022, 97.5% in 2023, 97.5% in 2024, 97.6% in 2025.

Any variable that assigns the same value to roughly 97 out of 100 observations carries almost no classifying power, regardless of how carefully each individual value was determined. This is a property of the distribution, not of the auditors.

Where the separation actually lives

The separation sits one level below the opinion, in the emphasis-of-matter paragraph. Of the 84 companies that held a clean opinion with a going-concern emphasis in 2024, 27 — 32.1% — were delisted or received a modified opinion in 2025. For companies without that paragraph, the comparable rate was 1.4%.

That is roughly a 23-fold difference in outcome between two groups that received the same audit opinion. By the standard of lift, it is an excellent flag.

The axis nobody quotes

Only 66 companies (2.5% of clean-opinion filers) carried the paragraph in FY2025. And here two properties move against each other.

The narrower a flag is drawn, the higher its precision and the more of the eventual failures fall outside it. The non-flagged group's rate is 1.4% — but that group is roughly thirty times larger (in 2024: 2,615 clean opinions, 84 flagged). A low rate over a large base is not a small number. Most of what goes wrong goes wrong outside the warning.

There is a third axis, and it is the one that decides usability: timing. An audit opinion is issued once a year, after the fiscal year has closed. The 27 failures happened in the following year. So the 23x lift belongs to a signal that opens its eyes once every twelve months.

Korea Parallel

This is precisely why RaymondsIndex evaluates deterioration risk (WP) on three separate axes rather than one — lift, coverage, and lead time — and refuses to trade one for another. A relational-risk signal that only reproduces what the annual audit already said has added nothing. The question we hold ourselves to is not "does this flag separate the groups" but "does it separate them before the statutory filing does, and does it reach the companies the statutory flag never touches."

The US data shows how differently the same label can be calibrated. Audit Analytics' twenty-year review reports 1,816 companies (24.3%) receiving going concern opinions in FY2022, and 2,853 (28.3%) at the FY2008 crisis peak. What is a 2.5% marker in Korea is a 24.3% marker in the US. The review also finds that since 2006, non-accelerated filers received going concern opinions at roughly ten times the rate of accelerated filers and eighty times that of large accelerated filers — the flag is reading firm size as much as firm distress. Korea's release points the same way: companies with under ₩100bn in assets had the lowest clean-opinion rate, 95.2%.

Academic frame

Gutierrez, Krupa, Minutti-Meza and Vulcheva (2020, Review of Accounting Studies 25(4)) tested exactly this question on 1996–2015 US data. Going concern opinions alone showed no better predictive power than a default model built from financial ratios — but there was imperfect overlap between the two, and adding GCOs raised the number of predicted defaults by 4.4% without increasing Type II errors. Their conclusion is the useful one: GCOs summarise a complex set of conditions that other predictors do not capture. Complementary, not superior. The methodological hazard underneath all of this was named decades ago by Zmijewski (1984, Journal of Accounting Research 22, 59–82), whose work on financial distress prediction showed how sampling and base rates distort exactly these comparisons.

What to do with this

August is half-year reporting season in Korea — statutory filings are due within 45 days of period end. If you read one thing in a review report this month, do not read the opinion line. Read whether anything was added below it.

Note on sources: two outlets reported the non-flagged comparison differently — one as "1.4 percentage points," the other as "1.4%." Only the latter is arithmetically consistent with 32.1%, and that is the reading used here.

#RaymondsRisk #RelationalRisk #CorporateGovernance #GoingConcern #BaseRates #AuditOpinion