Yang and An give the same coefficient two scales, an order of magnitude apart, leaving no headline number a trader can use. The abstract says a 1% increase in acute physical risk attention lifts fund excess returns by 24.675%. In the conclusion, the same focus earns "an annualized alpha of approximately 2.7%". Yet 2.7% matches the transition-risk coefficient of 2.746. The paper never reconciles the two magnitudes.
What the measure captures
Climate risk attention is extracted from text and divided three ways. Acute physical risk covers typhoons, floods and hail. Chronic physical risk covers shifts in temperature and precipitation, while transition risk includes carbon policy, new energy and technology change. The score comes from annual reports issued by the companies in each fund's portfolio, weighted by position size. The economic premise is simple: managers who price climate shocks and policy changes ahead of the market should earn alpha, with different payoffs across the three horizons.
The sample contains 435 open-end actively managed equity funds from East Money Choice, producing 2,801 fund-year observations from 2013 to 2023. Firm text comes from A-share annual reports scraped from 2003 to 2023, with 58,508 analyzable segments. Funds must have top-10 holdings equal to at least 40% of NAV and at least ten years of operating history. The return measure is Carhart four-factor alpha. A Fama-French three-factor version supplies a robustness check.
The baseline coefficients are 24.675 for acute risk (t = 3.34), 2.746 for transition risk (t = 2.58), 2.734 overall (t = 2.77), and 7.678 for chronic risk (t = 0.77, insignificant). Adjusted R-squared ranges from 0.118 to 0.121. The paper then uses Brinson attribution to divide excess return among allocation, selection and interaction, before closing the empirical analysis with a structured-fund split.
Units break the headline result
The acute physical risk variable has a mean of 0.00, a standard deviation of 0.01 and a maximum of 0.045. Treating "a 1% increase" as a 1.00 move in the regressor implies a change twenty-two times larger than the sample maximum. The coefficient's accompanying interpretation creates the problem.
The alternative reading fares no better. Multiplying a one-standard-deviation change in acute exposure, 0.01, by 24.675 produces 0.247 in whichever units alpha uses. Table 2 reports a Carhart_Alpha mean of 0.19 and a standard deviation of 1.32, without saying whether the values are percent, percentage points or decimals. The two readings fail in opposite directions, and neither is checkable.
Everything rests on this number.
The heterogeneity claim uses the same inflated percentage language, describing acute risk as roughly nine times transition risk.
Inside the dictionary
The dictionary does the heavy lifting. Yang and An begin a Word2Vec model with terms from Chinese climate policy documents and academic literature. They keep iterating until additional candidate terms expand the dictionary by less than 5%. The research team manually reviews the expanded list in several rounds, and ChatGPT adjudicates disputed terms. A climate term counts only when it appears within 50 words of an entry in a separate 1,823-item financial and operational risk dictionary. Sensitivity checks use windows of 25 and 100 words.
The finished dictionary contains 921 keywords: 180 acute physical, 228 chronic physical and 513 transition. A company's exposure equals the share of annual-report words assigned to each category. Word2Vec has already been used for this purpose in China. The paper's literature review credits Guo et al. (2024) with constructing a Chinese climate risk disclosure indicator from social responsibility reports through the same method. Yang and An contribute the move from companies to funds by aggregating firm scores with top-10 holdings weights.
Is this attention?
Sautner and coauthors, whose work the paper extends, observe a named party discussing a named subject. Their data come from English corporate conference call transcripts and capture what participants say about a firm's own climate exposure. Yang and An instead score the annual reports written by companies already held in a fund, then weight those scores by position size. The fund manager contributes none of the text.
The resulting variable is a portfolio composition indicator carrying an attention label. Table 1 makes the issue visible through the transition dictionary's most common terms: wind energy conversion, wind power generation, wind farm, wind turbine, wind power equipment, wind power station and wind power industry. "High-quality development" appears 463,850 times by itself. A high transition-attention score therefore identifies a fund holding wind and new-energy companies, alongside policy boilerplate. The 2.746 coefficient fits a sector tilt.
The practical implications section says acute-risk screening "could identify managers with genuine alpha-generating skills". Such a claim requires a measure of skill independent of portfolio holdings. The construction does not separate them.
Panel C changes the result
The paper offers three robustness checks. Replacing Carhart with the Fama-French three-factor model raises the acute coefficient to 37.512 (t = 6.03), a 52% increase over baseline. Controls for manager changes, scale, style shifts and fee ratios leave the acute coefficient at 24.901. The other three remain within a hundredth of baseline: overall 2.701, chronic 7.480 and transition 2.709.
Panel C instruments climate attention with its own lag on N = 2,361. The overall coefficient is 3.108, with t = 0.32. Acute falls to 15.205, with t = 0.19; chronic reaches 44.709, with t = 0.46; and transition rises to 7.030, with t = 0.66. Every t-statistic is under 0.7. The paper says Table 4 confirms that its main conclusions hold. Panels A and B support that statement. Panel C gives a different result: once lagged attention instruments for the contemporaneous regressor, every estimate loses significance.
The evidence therefore looks contemporaneous. It captures same-year co-movement between the subjects discussed by a fund's holdings and the returns on those holdings, just as the sector-tilt reading would predict.
Minus eight in the acute column
The mechanism analysis comes apart in its reported figures. For acute risk, the Brinson allocation coefficient is -2.6516, selection is -9.3565 (t = -4.79), and interaction is +4.0293. These three components decompose a single excess return and sum to -7.98. The authors interpret the positive interaction coefficient as evidence of a contrarian strategy, in which funds buy oversold companies in distressed sectors and capture the rebound. Maybe. The +4.0293 interaction term still must overcome a selection effect more than twice as large at -9.3565.
The remaining columns sum to positive values. Chronic totals +9.27, comprising allocation of +14.0769, selection of -7.9896 and interaction of +3.1841. Transition totals +0.28, from allocation of +0.6654, selection of -0.2472 and interaction of -0.1411. Across every column, selection is negative and allocation is positive. The authors call this an asset allocation-stock selection paradox and connect it to closet indexing, a fair and interesting interpretation. The same pattern also fits a variable that records sector exposure more readily than manager judgement.
The heterogeneity test deserves equal care. Structured funds are those that did not change manager after 2020, while funds that did are classified as non-structured. Manager turnover alone defines the split. The non-structured acute coefficient is 31.3270 (t = 3.13), compared with 8.5940 (t = 0.60) for structured funds. Those estimates describe funds with recent manager changes.
What remains usable
Timing imposes a structural constraint. Firm annual reports for year t arrive in year t+1, as does the fund's holdings disclosure. The paper explicitly assumes that a fund chooses its portfolio at the beginning of the year and leaves it unchanged. The assumption works within an attribution exercise and offers nothing for a live screen. Requiring a ten-year operating history also removes failed and closed funds by construction. The paper gives the filter and its rationale, sufficient time-series data, without naming that consequence.
The abstract acknowledges the annual-frequency limitation. The conclusion also discusses the supply-chain gap, the static Brinson framework and generalizability beyond Chinese equity funds. Those concessions leave the paper's decisive problems untouched: the units attached to its headline coefficient and the evidence in Panel C.
We could not test any of these results. Doing so requires Chinese A-share annual report text and fund-level holdings from East Money Choice, and we have neither. Using US filings and US funds would produce a different study.
The dictionary remains the contribution worth retaining: 921 Chinese climate terms, divided three ways and paired with a documented context filter. Table 1 also displays its noise. "wind speed" appears in the acute list 1,985 times and in the chronic list 31,682 times, while "high-quality development" dominates the transition column with 463,850 appearances. The alpha claim layered onto that dictionary still needs a predictive test. Its lagged-instrument specification already produces no significant result, with all four t-statistics under 0.7. Forming portfolios from lagged, publication-date-aligned exposure and reporting the long-short spread and turnover would turn the claim into a testable signal.