An eleven-month ESG null the panel cannot support

Eleven months of returns cannot establish whether Refinitiv ESG scores are priced in Southeast Asia, even within the short horizon the authors claim. The paper's arithmetic gives a portfolio manager another reason to put little weight on its null.

What did Doblas and co-authors test?

The premise is familiar: markets that reward sustainability should deliver higher returns to firms with better or improving ESG scores. The authors offer two channels. Scores may signal lower agency costs and regulatory risk, or ESG-minded investors may bid up the shares. Doblas, Lagaras, Regadio, Panerio, Ekundayo and Salindo study 327 listed Southeast Asian firms with full Refinitiv pillar coverage. Their regressions pair monthly USD returns with same-month ESG scores from February to December 2023. Refinitiv's sector-weighted scores range from 0 to 100. The composite has a mean of 54.8 and a pooled SD of 8.5.

The paper uses three specifications. First comes a panel with firm and time fixed effects and firm-clustered errors, estimated separately for the composite and the three pillars jointly. Quantile regressions follow at the 25th, 50th and 75th percentiles. These use only the composite, replacing firm effects with sector and industry dummies. The authors' final check is pooled OLS with those same dummies and Newey-West errors at six lags. They winsorize returns and ESG at 1/99 and z-score ESG.

The composite produces a null throughout. Its fixed-effects coefficient is β = -0.00010 (p =.868), with within R² of 0.00003. The estimate at the 25th quantile is -0.00004 (p =.795). Environmental supplies the lone significant result in the pillar fixed-effects model: β = 0.0013 at p =.014. That significance disappears in the HAC regression. The authors conclude that short-run Southeast Asian stock returns do not embed ESG practices.

The panel has more rows than it can hold

327 firms observed for 11 months allow at most 3,597 firm-months. The reported count is 3,778 observations, an excess of 181 over a balanced panel. Each of the paper's three filtering steps can only remove rows. Duplicate firm-months could account for the gap; so could a twelfth month. An implementer has no grounds to choose between them.

The tables also disagree with each other. The fixed-effects composite model has 3,778 rows, against 3,776 for the pillar model. In the HAC table, Model 1 (one ESG regressor) reports N = 3,776 and F(79, 3696). Model 2 (three pillars) reports N = 3,778 and F(77, 3700). Giving the single-regressor model two more numerator degrees of freedom than the three-regressor model suggests the columns were swapped.

No down months

The dependent variable is defined as the percentage change in adjusted USD close. Yet its raw distribution runs from a minimum of 0.002 to a maximum of 2,953.732. The median is 0.736, the mean 9.281, and skewness 19.47. For 327 stocks during 2023, every firm-month is nonnegative.

Those figures resemble a price-level distribution. If the dependent variable is a level, firm fixed effects absorb that level; the composite null then describes price noise around firm means. Winsorizing at roughly 0.007 and 104.22 cuts the tails while leaving the variable's meaning unchanged. Before writing any code, a replicator must guess what the authors calculated.

Environmental loses its fixed effects too

The authors interpret the disappearance of Environmental's significance as "implying that the initial significance was perhaps driven by serial correlation." But the fixed-effects errors were already clustered by firm. The paper itself says this corrects within-firm serial correlation and heteroskedasticity. Moving to the HAC regression changes much more than the covariance estimator: firm and time effects disappear, replaced by sector and industry dummies. The coefficient consequently draws mostly on cross-sectional differences in score levels.

The HAC pillar coefficients appear only as "n.s." at p >.10. We therefore cannot distinguish a smaller point estimate from a larger standard error. And a six-lag Newey-West window covers more than half an 11-month series.

Identification is also a problem on the fixed-effects side. The paper says ESG disclosures are annual and Refinitiv revises scores as information arrives. A firm would therefore have at most one change in 11 months if updates follow annual disclosures. We did not find any within-firm standard deviation for the scores in the paper. It reports pooled figures instead, including the composite's 8.5. The pillar model's within R² of 0.0008 is consistent with slopes estimated from very little movement.

Repeated digits across estimators

For the 75th quantile, the table gives ">.10" without a coefficient. The 25th-quantile and HAC composite estimates are identical to every printed digit: -0.00004 at p =.795. One comes from quantile regression, the other from pooled OLS. The normality table has another repetition: its kurtosis and p-value columns are equal (0.991, 0.987, 0.993, 0.989). Both patterns suggest copied cells.

What remains for a trader after 11 months?

Very little in either direction.

All the regressions are contemporaneous: ESG in month t is paired with return in month t. None uses a lagged score or a portfolio sort, and none includes a time-varying beta or momentum control. The paper depends on firm fixed effects to absorb size. Its quantile and HAC specifications then remove those effects. The authors acknowledge in their limitations that they do not examine dynamic or forward-looking effects.

Taken at face value, a near-zero coefficient describes same-month co-movement over 11 months. Whether a high-minus-low ESG sort earns a premium after costs remains unanswered. The conclusion goes further, characterizing ESG in the region as merely a long-term indicator of the underlying institution's quality. Yet the authors defer the long-horizon test to future work. They assert a long-term role while explicitly leaving it untested. An 11-month panel cannot substantiate a long-horizon claim in either direction.

We could not rerun the study. We have neither Southeast Asian equity prices with the historical listing universe nor monthly point-in-time Refinitiv pillar histories. Financial statements and news cannot replace the latter.

A multi-year panel would settle the question if it lagged scores a month, used a return series that goes negative, and retained fixed effects in the HAC run. If the composite were still flat, I would accept the null as real. This paper has yet to show that.