A first social disclosure in an SEC filing is a usable risk flag, though the paper cannot establish its economic size. Idiosyncratic volatility rises with that first mention. The equivalent appearance in a stand-alone sustainability report produces nothing in the pooled and topic-level regressions. This channel gap carries the paper through topic splits, downside residuals and propensity-score matching. One matched human-capital sample breaks the pattern. Rebuilding the central result from the published description remains impossible.
Why the channel matters
Hoepner, Korca, Schiemann and Schneider attribute the difference to legal exposure. Once management places a social issue in Item 1A risk factors, its assessment becomes actionable under Rule 10b-5. The market may then read a first mention as evidence that a topic has moved from corporate-responsibility narrative into latent liability. Voluntary channels, in the authors' words, "lack the other two elements (being highly regulated and containing financially material information)". Binding materiality and enforcement apply only to the regulated channel, which is where the authors expect surprises to enter prices.
The sample contains 755 US companies from the S&P 1,500 over 2011 to 2015, producing 2,982 firm-years. Datamaran supplies the disclosure data. Its machine-learning text extraction identifies 109 ESG items in SEC filings, financial reports, and stand-alone sustainability reports. The authors manually assign them to 20 human capital items, 11 product liability items and 15 stakeholder engagement items. For every topic-channel-year, a binary variable records whether at least one item appeared.
Annual idiosyncratic risk follows Fu (2009). It comes from firm-level daily Fama-French three-factor regressions, while the downside measure retains residuals from negative-market days only. Mean idiosyncratic risk is 0.014. Its 99th percentile is 0.034.
Table 3 supplies the main result. First-time SEC disclosure has a coefficient of 0.0531 (SE 0.0136) in the all-channel model and 0.0542 (SE 0.0138) when the SEC channel stands alone. In the same aggregate model, first-time sustainability-report disclosure is 0.0044 (SE 0.0193), while first-time financial-report disclosure is -0.0226 (SE 0.0242). Both estimates are insignificant, matching the authors' hypotheses.
Continued voluntary disclosure moves in the other direction. Sustainability reports load at -0.0338 and financial reports at -0.0542 in the all-channel model, both p<0.1. Their channel-specific estimates sharpen to -0.0422 and -0.0592 at p<0.05.
The SEC result persists across topics. First-time disclosure loads at 0.2494 for human capital, 0.2409 for product liabilities and 0.0953 for stakeholder engagement. The downside estimates are 0.2392, 0.2175 and 0.1414. After one-to-one matching without replacement at a caliper of 0.03, the coefficients are 0.3726 (n=497), 0.2670 (n=494) and 0.1488 (n=601), all p<0.01. The exception appears in the matched human-capital sample, where a sustainability-report first mention also loads positively, at 0.1823 and p<0.01. It is the only voluntary first mention that resembles the SEC result.
Financial-report estimates at topic level also conflict with H1c. First-time disclosure is negative and marginally significant (p<0.1) in four of six models. For idiosyncratic risk, the coefficients are -0.0345 for human capital, -0.0451 for product liabilities and -0.0456 for stakeholder engagement. Stakeholder engagement produces -0.2494 on downside risk. The hypothesis predicted a null. The authors therefore describe the voluntary evidence cautiously: they "find only weak evidence of such a relation for social disclosure via the channels of sustainability reports and financial reports."
The result worth watching
The interaction carries more trading relevance than the standalone estimates. When a topic first appears in both the 10-K and sustainability report during the same year, the estimated effects compound. Human capital adds an interaction of 0.5189 (SE 0.1105) to a main effect of 0.2393, roughly tripling the SEC-only figure. Product liabilities add 0.3600 (SE 0.1497, p<0.05) to 0.2587, about 1.4 times. Stakeholder engagement adds 0.2559 (SE 0.0842) to 0.1187.
Extra detail amplifies rather than dampens.
Earlier disclosure does dampen the response. The interaction between first-time SEC and continued financial-report disclosure is -0.2063 (p<0.1) for product liabilities and -0.2483 (p<0.05) for stakeholder engagement. Pairing first-time SEC disclosure with continued sustainability-report disclosure gives -0.0653 (p<0.1) for stakeholder engagement. Continued SEC human-capital disclosure also works against H2a, loading positively at 0.0568 on idiosyncratic risk and 0.0822 on downside risk. The authors acknowledge that support for H2a through H2c shifts across models.
Which "first-time" variable was estimated?
Two incompatible definitions govern the paper's central variable. Section 4.1 assigns one to the first-time dummy when a company discloses during the current year and "not in any of the previous years in our sample". Table 1 instead assigns one when the company reports through the relevant channel on at least one social topic absent from the previous year's report. A never-before mention and a return after one silent year generate different variables. The continued dummies contain the same ambiguity.
Datamaran coverage begins in 2010, which is burned as the base year. Consider a firm that disclosed in 2010, stayed silent in 2012 and resumed in 2013. One definition treats it as a first-time discloser; the other treats it as lapsed. A footnote acknowledges the left censoring. As far as I can see, the paper never identifies which coding rule generated Table 3.
Units remain unstated
Size loads at -0.1620 and SDROA at 3.0477, even though the dependent variable's 99th percentile is 0.034. Some rescaling occurs between the descriptive statistics and regression tables, yet I did not find the factor. Table 1 adds another conflict by defining Id_risk as the "squared residual of three factor model". The methods section calls it the standard deviation of residuals across one year. Without a settled definition and scale, 0.0531 has no interpretable economic magnitude.
The filing date disappears inside the risk window
The proposed mechanism says repricing "manifests as a short-term spike in idiosyncratic volatility." Measurement occurs over a calendar year: an annual disclosure dummy is regressed on the year's residual standard deviation. The filing date lies inside that outcome window. This setup cannot distinguish volatility caused by the filing from volatility that encouraged management to add risk-factor language.
The tables describe relations. Stronger causal wording enters the results and conclusion, where "first-time SEC disclosure increases idiosyncratic risk" and disclosure location and timing "fundamentally alter firm-specific risk profiles". Neither statement rests on an event window, and the paper conducts no event study. Matching changes the functional-form comparison. A latent shock could still drive both the disclosure language and volatility.
The interaction raises an additional arithmetic concern. First-time SEC human-capital disclosure appears in 3.0% of firm-years, compared with 5.0% for sustainability-report human-capital first mentions. Under independence, their joint cell would contain about 0.15% of 2,982, or four or five observations. A standard error of 0.1105 seems too tight for such a cell, suggesting that first mentions cluster across channels. I did not find the joint count.
Useful as a monitor
A first appearance of human capital or product liability language in a 10-K can flag a name for wider risk bands. The signal looks strongest when a sustainability report introduces the same topic at the same time. It offers nothing as alpha, a claim the paper does not make: there is no portfolio or cost analysis, and the conclusion concerns disclosure policy.
Nine channel-topic combinations across two risk measures also create a substantial testing burden. I found no multiple-testing adjustment. SEC coefficients reach p<0.01 in every model except stakeholder engagement on idiosyncratic risk in Table 4. The table reports 0.0953 at p<0.05, although the text claims p<0.01 for all six topic models. This is a third reporting inconsistency. The voluntary-channel estimates remain the uneven part of the evidence.
We could not test the result. It depends on Datamaran's channel-specific classification, which we do not license. We also lack a point-in-time corpus of filings and sustainability reports from 2011 to 2015, so we cannot reconstruct the first-mention labels. Our text coverage begins around 2020. By the authors' own reason for ending the sample at 2015, that belongs to the wrong regime.
A filing-frequency version would change my view if it measured residual volatility in a window beginning on the 10-K date and stated the dependent variable's scaling.