Does extreme climate concern drive equity premiums? Evidence from China

The paper investigates whether heightened concern about extreme climate conditions, as reflected in news media coverage, predicts equity risk premiums in China. Motivated by growing evidence that climate risks affect asset prices and by regulators’ emphasis on climate-related financial stability risks, the authors construct a news-based Extreme Climate Concern indicator (ECC) and test its ability to forecast stock market returns. The study aims to determine: (1) whether ECC predicts equity premiums; (2) how its predictability compares with conventional confidence and macroeconomic predictors and with existing climate risk indices; (3) whether predictability differs before and after the Paris Agreement; (4) whether predictability varies across industries and seasonal high/low climate concern periods; and (5) whether ECC delivers economic value in asset allocation. The importance lies in identifying a climate-news-driven predictor with incremental information content for pricing in a major emerging market.

The study builds on literature showing climate risk’s pricing implications. Pástor et al. (2022) and Ardia et al. (2022) document that green stocks outperform brown stocks when climate concerns rise, indicating markets react to climate news. Barnett (2023) shows climate model uncertainty amplifies the price of climate risk. In fixed income, Huynh and Xia (2021) find bonds with higher climate news betas have lower future returns, suggesting priced climate risk. Schlenker and Taylor (2021) show weather forecasts and warming trends affect derivative prices. Sautner et al. (2023b) find firms with higher climate risk exposure earn risk premia ex ante, with prices adjusting as risk materializes. Institutional investors value climate risk disclosure (Krueger et al. 2020; Ilhan et al. 2023). Extreme heat reduces firm revenues and operating income (Pankratz et al. 2023). Prior China-focused indices capture climate attention/physical risk (e.g., CPR, CAI: Guo et al. 2023; CRI: Zhang et al. 2023a; CRP: Ji et al. 2022). The paper complements this work by proposing a focused extreme-climate news-based indicator (ECC) and assessing its predictive and economic value in China.

Data and ECC construction: News data come from the CSMAR market information module (financial news unit), yielding 7,816,916 articles from 2008-01-01 to 2021-12-31. The authors compile a dictionary of 58 extreme climate keywords (e.g., gale, rainstorm, blizzard, tornado, hail, typhoon, strong cold wave, sandstorm, fog, prolonged drought, high wind). For each month, they count keyword mentions (n_t) and define the ECC using a log-difference to improve stationarity: ECC_t = ln(1 + n_{t+1}) − ln(1 + n_t). They also document ECC’s time-series evolution (2008:01–2022:12), noting a sharp rise post-Paris Agreement (2016-11). Additional data include market returns (China A-share composite including dividends) minus the 3-month Treasury bill rate as excess return R^m, confidence indicators (BCI, CLI, CCI, CICSI, CPI), and macro predictors (FSI, RV, RSKEW, RKURT, DP, DY, DE, PPI) from WIND/OECD. Descriptive statistics and correlations are reported (Tables 1–2). In-sample predictive models: Univariate predictive regressions R^m_{t+1} = α + β Z_t + ε_{t+1} where Z_t is ECC or any of the 13 alternative predictors; and bivariate regressions R^m_{t+1} = α + β ECC_t + ς Z^k_t + ε_{t+1} to test ECC’s incremental predictability over each alternative predictor (Eqs. 2–3). They also compare ECC with four climate risk indices: CPR, CAI (Guo et al. 2023), CRI (Zhang et al. 2023a), and CRP (Ji et al. 2022), via R_t = α + β ECC_t + γ ClimateRisk_t + ε_t (Eq. 4). They evaluate industry-level predictability using CSRC’s 16-industry classification. Before/after Paris Agreement: They split the sample to assess predictability over several horizons (1–6, 9 months) before and after the Paris Agreement’s implementation in Nov 2016 (Table 6). Out-of-sample (OOS) forecasts: Recursive estimation following Welch and Goyal (2008) with initial in-sample period 2008:01–2013:12 and evaluation 2014:01–2022:12. The OOS forecast uses R_{t+1} = α_t + β_t ECC_t; they compute OOS R^2_os and MSFE-adjusted statistics relative to the historical mean benchmark (Eqs. 6–9). They also examine an alternative OOS window (initial 2008:01–2016:12; evaluation 2017:01–2022:12; Table 15). Economic value: Mean-variance portfolio allocation using OOS forecasts to set the weight in risky assets w_t+1 = R̂_{t+1}/(γ σ̂^2_{t+1}), with σ̂^2 estimated from a 60-month rolling window. They compute certainty equivalent return (CER) gains and Sharpe ratios for risk aversion γ ∈ {1,3,5} (Eqs. 10–12; Table 10), comparing ECC-based strategies with those using alternative predictors and with bivariate combinations. High/low concern seasons: Using a two-state model (Huang et al. 2017), they define April–October as high-concern and November–March as low-concern periods and estimate state-dependent predictability in-sample and out-of-sample (Eqs. 16–17; Table 14). Alternative ECC via 3PRF: They construct a consistent ECC factor (ECC_3PRF) using the three-pass regression filter (Kelly and Pruitt, 2015). Step 1: time-series regressions of lagged keyword attention E_{it−1} on future market return R_t to obtain β_i sensitivities (Eq. 13). Step 2: cross-sectional regressions E_{it} = φ_i + ECC_t β_i + u_{it} to estimate ECC_t (Eq. 14). They provide the compact estimator (Eq. 15) and test ECC_3PRF’s in- and out-of-sample predictive power over horizons (Table 11). They further test ECC across industries and across the Shanghai and Shenzhen markets (Table 13).

- ECC significantly and negatively predicts next-month market excess returns. In univariate in-sample regressions, β = −1.815% (t = −2.693), R^2 = 4.012% (Table 3). - ECC’s predictive power is robust to including alternative predictors. In bivariate models with each of 13 predictors, ECC remains negative and significant (β between −1.956 and −1.709). R^2 rises notably when combined with BCI (R^2 = 8.788%), CPI (10.885%), and PPI (10.365%), indicating complementarity (Table 3). - Compared to existing climate indices (CPR, CAI, CRI, CRP), ECC remains significant while the others are mostly insignificant in joint regressions, and combined models produce higher R^2 than ECC alone (Table 4), evidencing incremental information. - Industry-level results show significant negative predictability for many sectors: Water and Electricity (β = −1.961***, R^2 = 4.635%), Architecture (−2.171***, 4.670%), IT (−2.261***, 4.541%), Real estate (−2.075***, 4.983%), and Scientific Research (−2.156***, 4.688%). Effects are insignificant for Mining, Finance, Business, and Culture (Table 5). - Before vs. after Paris Agreement: In-sample predictability is stronger and persists at longer horizons after the Agreement. Post-Paris, 1–6 month horizons are significant with R^2 between 4.048% and 7.003% (Table 6). Pre-Paris, only the 1-month horizon is significant (R^2 = 3.727%). OOS results similarly improve post-Paris (Table 9): after Paris, R^2_os ranges from ≈2.63% (1-month) to ≈5–6% (2–6 months), whereas pre-Paris long-horizon OOS R^2_os are negative or insignificant. - Out-of-sample predictability: Univariate ECC achieves R^2_os = 2.151% (MSFE-adj = 2.146, p = 0.017), exceeding the historical mean. Among comparators, BCI (1.485%), CLI (1.712%), and DY (3.334%) perform well. Combining ECC with BCI/CLI/DY yields larger R^2_os: ECC&BCI 4.062%, ECC&CLI 3.834%, ECC&DY 4.854% (Table 8). Results hold under an alternative OOS window (Table 15). - High vs. low concern seasons: ECC’s predictive power is stronger in low-concern months (Nov–Mar). In-sample, low-concern β estimates are more negative and significant (e.g., −2.491*** for value-weighted). OOS R^2_os are larger in low-concern than high-concern periods (Table 14), consistent with limited investor attention. - Economic value: ECC-based timing strategies deliver positive CER gains and Sharpe ratios across risk aversion levels, and combining ECC with BCI/CLI/DY further increases CER and Sharpe (Table 10). For example, ECC&BCI attains CER gains up to 3.489% (γ = 1) and Sharpe 0.322. - ECC_3PRF robustness: The 3PRF-based consistent ECC shows significant in-sample R^2 over multiple horizons (e.g., 1-month R^2 = 2.630%; quarterly 5.710%; semiannual 5.296%; annual 1.643%) and strong OOS R^2_os from 1.633% to 6.559% (Table 11), reinforcing robustness. - Across markets: ECC predicts returns in both Shanghai (in-sample R^2 = 3.403%; OOS R^2_os = 2.562%) and Shenzhen (in-sample R^2 = 4.398%; OOS R^2_os = 1.719%), with BCI also predictive (Table 13).

The findings confirm that extreme climate concern, as captured by news coverage, contains forward-looking information about equity risk premiums in China. Elevated ECC is associated with lower subsequent market returns, consistent with channels whereby extreme weather risk impairs macroeconomic activity, raises operational costs, shifts demand, and heightens risk perceptions; and with behavioral channels where climate-related news increases investor anxiety, leading to selling pressure. Stronger predictability after the Paris Agreement suggests that policy salience, intensified climate action, and improved climate awareness have increased the transmission of climate information into asset prices. The complementarity between ECC and traditional confidence and macro indicators implies ECC captures distinct information not subsumed by economic fundamentals or broad sentiment. The stronger predictability during low-attention seasons supports limited attention: when climate attention is lower, markets underreact more to climate signals, improving predictability. Industry heterogeneity aligns with differing exposure to climate risk (e.g., utilities, construction, real estate, IT, and scientific research appear more sensitive). The demonstrated economic value indicates practical relevance for asset allocation and risk management. Overall, the results advance climate-finance integration in pricing and forecasting within an emerging market context.

This study introduces a news-based Extreme Climate Concern indicator (ECC) for China and shows it is a powerful, negative predictor of equity premiums in-sample and out-of-sample. ECC outperforms most conventional confidence and macro predictors and provides incremental information relative to existing climate indices. Predictability strengthens after the Paris Agreement and is higher during low-attention seasons, consistent with policy effects and limited investor attention. ECC-based strategies yield appreciable CER gains and Sharpe ratios, and robustness holds across industries, exchanges (Shanghai/Shenzhen), alternative OOS windows, and with a consistent factor constructed via the 3PRF method (ECC_3PRF). Future research could extend to firm-level cross-sections and international markets, integrate higher-frequency news and alternative media sources, explore causal identification of climate-news shocks, and assess interactions with transition policies and physical risk measures.

The paper does not center on causal identification; results are predictive associations. The ECC relies on a news-based keyword dictionary and CSMAR coverage, which may introduce measurement noise or media bias. Some comparative climate indices (CPR, CAI, CRI, CRP) have shorter sample ranges due to data availability, potentially affecting comparisons. The seasonal high/low concern classification is based on Chinese climatic patterns and may not generalize across regions. Findings are specific to China’s A-share market and the 2008–2022 period.