A one-SD increase in the Huazheng ESG score accompanied about 0.54 points of additional cumulative abnormal return through the tenth session after the event day. This is the paper's [0,10] estimate for 5,041 A-shares following the 7 October 2023 Hamas attack. The regression holds eight firm-level controls and 21 industry dummies fixed. An ESG screen has some support here. The paper takes the trading case no further.

What trade follows from ESG and ROA?

He, Gong, Ng and Lau build their argument around the ownership of Chinese equities. Citing Gao and Yang (2018), they put individual investors' share of the Chinese stock market at more than 60%. Their premise is that retail holders respond to bad geopolitical news by selling in herds.

The selling should hit firms differently. Highly profitable companies, bid up before the shock, should suffer larger falls. Strong ESG profiles should offer some protection through stakeholder goodwill.

The sample contains 5,041 A-shares listed in Shanghai, Shenzhen and Beijing. For each stock, the authors estimate two expected-return models over 160 trading days, from 9 February to 14 September 2023. A market model uses the stock's own exchange index. A Fama-French three-factor model is also constructed for each exchange. Abnormal returns accumulate from the event day, 9 October 2023, the first trading session after the attack.

The authors then run cross-sectional regressions. They take each firm's cumulative abnormal return (CAR) over [0,1], [0,5] or [0,10] and regress it on one of three variables:

All regressions contain 21 industry dummies. The authors run versions with and without eight firm-level controls and cluster standard errors by city. For a trader, the implied response to a geopolitical shock is to favour high-ESG stocks and reduce exposure to high-ROA stocks.

Losses relative to the index

Under the market model, CAAR, the average CAR across stocks, reaches -0.0192 over [0,10], with t = -17.07. The three-factor estimate for that window is -0.0085. SMB and HML therefore remove more than half the measured loss. For [0,5], the respective estimates are -0.0059 and -0.0033.

Equal-weighted stock averages are being compared with cap-weighted exchange indices. Adding a size factor in the three-factor model reduces the [0,10] loss from -0.0192 to -0.0085. This is consistent with small caps underperforming the cap-weighted indices. Both estimates measure performance relative to a benchmark. We did not find the indices' own returns for the window reported, leaving readers unable to recover the outright loss on a long book.

The t-statistics warrant a large discount. The 5,041 stocks all share the same event date, and their abnormal returns are correlated. A cross-sectional t-test treats those returns as independent draws.

We did not find a dependence adjustment.

There is another problem with the date. China's exchanges reopened after the National Day holiday on 9 October. The event-day (day 0) return therefore takes in a week of holiday news. Only the component each stock shares with its exchange index gets netted out. We did not find a discussion of the holiday.

The authors acknowledge that their method cannot "completely isolate the effects of concurrent events." They say a narrower sample period helps mitigate this problem and concede that the method may introduce slight estimation biases. Narrowing the window offers no help when a concurrent shock arrives on day 0.

The size of the ESG association

The controlled ESG coefficient for CAR[0,10] is 0.0130 (SE 0.0030). Multiplying it by the score's SD of 0.412 gives about 0.54 percentage points of ten-day abnormal return for a one-SD higher score. CAR[0,10] has a cross-sectional SD of 0.084. The ESG association thus amounts to about 6% of one SD of CAR, with adjusted R-squared of 0.0142.

Controls make a substantial difference. Omitting them lowers the [0,10] coefficient to 0.0060, significant only at 10%, or about 0.25 points. The IV estimate is 0.0104.

The 0.54 measures a fitted difference in conditional means. It comes from one event, in-sample and before costs, with eight controls and industry held fixed. The authors describe high-ESG firms as having "more stable stock returns" and "greater stability". Every outcome in Tables 3 to 8, however, measures a CAR level. None measures volatility or dispersion. Higher average CAR is the claim the evidence supports; stability is a different claim.

ROA produces the more surprising result. With controls, its CAR[0,10] coefficient is -0.0922, equivalent to about -0.55 points per ROA SD of 0.060. For [0,5], the IV estimate of -0.2196 is more than double the OLS estimate of -0.1011. The authors attribute this to herding into widely held winners. Their regressions contain no retail-ownership or turnover data.

For ROA on CAR[0,1], adding eight controls lowers adjusted R-squared from 0.0985 to 0.0313. Both regressions have 5,041 observations. Such a decline is difficult to reconcile with a nested OLS fit.

Identical estimates across the three-factor horizons

The three-factor replication has nine columns covering three horizons. Each [0,1] column repeats exactly in its [0,5] and [0,10] counterparts. All ROA columns report -0.0562 (0.0161), adjusted R-squared of 0.0540 and F of 7.3354. All ESG columns report 0.0072 (0.0013), adjusted R-squared of 0.0513 and F of 11.0025. Yet the underlying three-factor CAARs change with the horizon, from -0.0032 to -0.0085.

The authors say the signs "remain consistent" and describe the table as "demonstrating the consistency of these findings with the results based on the CAR calculated using the MM". Identical columns can provide evidence for at most one horizon. The claim of consistent findings rests on those repeated columns. We raised a similar unresolved table problem in an earlier event study of credit-rating announcements.

The Bloomberg check supplies separate evidence. Across 615 firms, ESG coefficients range from 0.0003 to 0.0010 and are positive and significant at 1% in all six specifications. A second rater therefore corroborates the sign. Bloomberg's scale differs from the normalised Huazheng score, so the magnitudes cannot be compared. Model 31's F of 1.3527 is also insignificant.

One shock date leaves the screen untested

A shared event date supplies one observation of the market's mood. Cross-firm ESG differences remain observational. In defending their instruments, the authors report Hansen J p-values ranging from 0.1834 to 0.9043 and an ESG first-stage Kleibergen-Paap Wald F of 5869.18. They say the IVs confirm that the regressors are uncorrelated with the error term.

The overidentification test cannot detect two instruments invalid in the same direction. Within a single cross-section, the industry-mean instrument is also near-collinear with the 21 industry dummies. The ESG first stage reports a constant of -23.5481 for a variable whose SD is 0.412. We also did not find an as-of date for the Huazheng scores.

We could not backtest the result. Our data contains neither a broad A-share universe nor historical Chinese ESG scores. Using US stocks instead would abandon the retail-ownership premise on which the paper depends.

A faithful trading test would require rules the paper leaves unspecified. It reports regressions rather than a portfolio, so a formation date, weights, costs and A-share price limits would all need to be set. We would be persuaded by the 0.0130 surviving several unrelated shocks and placebo dates outside Golden Week.