Does carbon emission rights trading foster energy-saving technological innovation? Evidence from listed manufacturing firms in China

Summary

China’s green transition is advancing, yet fossil fuels remain dominant, underscoring the importance of energy-saving technological innovation (ESTI) for low-carbon development. Existing studies mainly focus on the impact of carbon emissions trading (CET) on broad green innovation, while evidence on CET-induced ESTI remains limited. Using 3.09 million patent texts from Chinese listed manufacturing firms during 2008–2023, we employ a large language model (LLM) to identify firm-level ESTI and estimate the effect of CET through a staggered difference-in-differences (DIDs) approach. Results show that CET significantly promotes ESTI, and the findings remain robust across multiple tests. Mechanism analysis reveals three channels: increased carbon-risk exposure, substitution between government subsidies and market incentives, and correction of market inefficiencies. The effect is stronger among firms with lower pollution intensity and those without greenwashing behavior. These findings provide micro-level evidence on how environmental regulation fosters ESTI and offer implications for carbon-market and subsidy design.

Graphical abstract

Highlights

• •Firm-level data: CET boosts energy-saving tech innovation (ESTI) in China
• •We use large language models to identify firms’ ESTI from patent texts
• •Channels: carbon-risk exposure, subsidy substitution, market compensation
• •Carbon trading works better in low-pollution and non-greenwashing firms

Environmental science; Pollution; Global Carbon Cycle; Manufacturing

Introduction

At the 75th UN General Assembly, China pledged to peak carbon dioxide emissions before 2030 and achieve carbon neutrality before 2060, embedding this commitment in its national strategy and accelerating the green transition.¹ Data released by the National Bureau of Statistics on China’s total energy consumption and consumption mix from 2008 to 2023 show that, amid steadily rising energy demand, the share of natural gas and clean energy sources such as wind and solar power has grown rapidly, underscoring strong momentum in the low-carbon energy transition. Nevertheless, traditional fossil fuels, particularly coal and oil, remain the dominant sources of energy consumption. In 2023, fossil fuels such as coal and oil accounted for 73.6% of China’s total energy consumption (see Figure S1). Fossil fuels are therefore the primary source of China’s carbon emissions, and reducing their use constitutes an essential pathway toward low-carbon development.² Energy-saving technologies innovation (ESTI) play an increasingly pivotal role in this process. By improving energy efficiency and reducing reliance on conventional energy sources, such technologies demonstrate remarkable potential for cutting carbon emissions. Accelerating their research, development, and deployment has therefore become a key strategic measure for advancing China’s dual carbon goals.

Against the backdrop of deepening commitment to sustainable development, enterprises pursuing energy-saving transitions commonly face multiple constraints, including lengthy cost-recovery periods, intensifying market competition, and insufficient technical expertise.^3,4 Effectively advancing this transition hinges on well-designed market mechanisms to alleviate these constraints, among which the carbon emissions trading (CET) system plays a particularly important role. Since 2013, China has launched seven regional carbon trading pilots in Beijing, Shanghai, and other areas, followed by the official launch of a unified national carbon market in 2021, which initially covered the power generation sector. Unlike conventional command-and-control approaches to emissions reduction, CET is a market-based environmental regulatory instrument: By setting carbon emission allowances for enterprises and permitting trading within these allowances, it uses price signals to incentivize firms to proactively reduce carbon emissions and improve energy efficiency.⁵ The Porter Hypothesis posits that well-designed environmental regulation, far from undermining corporate competitiveness, can stimulate technological progress and managerial innovation, thereby partially offsetting or even outweighing the cost burdens associated with environmental protection.^6,7,8 As a flexible and effective emissions reduction instrument, CET aligns intrinsically with the Porter Hypothesis: It leverages market-based mechanisms to motivate corporate energy-saving efforts and rewards such behavior with additional market returns, thereby enhancing firms’ competitive advantage.

Current research on CET and energy technology innovation has become relatively extensive.^9,10 While these studies provide theoretical foundations and empirical evidence for understanding how CET drives ESTI, several limitations remain: In terms of research focus, existing literature has predominantly treated green technology innovation as an aggregate category, paying insufficient attention to ESTI, which is a specific type of innovation that directly enhances energy utilization efficiency. In terms of measurement methodology, existing studies have either relied on energy efficiency indicators at the prefecture-city level or used electricity-saving patents alone as a proxy, making it difficult to comprehensively characterize the multidimensional ESTI activities of firms at the micro level. In terms of mechanism identification, existing studies have provided insufficient characterization of the internal channels through which CET influences corporate innovation decisions, and the relevant black box has yet to be adequately unpacked.

Building on the limitations identified above, we propose the following two core research questions:

• •Has the implementation of CET effectively promoted the development of ESTI in manufacturing firms?
• •Through what channels can the implementation of CET contribute to ESTI, and what are the underlying mechanisms at work?

To address these questions, we draw on 3.09 million patent text records from Chinese listed manufacturing firms over the period 2008 to 2023, and employ the Enhanced Representation through Knowledge Integration (ERNIE) large language model (LLM) to measure firms’ level of ESTI. We then adopt a difference-in-difference (DID) model to investigate the effects of CET on ESTI and to identify the underlying mechanisms. The findings indicate that CET promotes ESTI in manufacturing firms by intensifying carbon risk exposure, exerting the incentive-substitution effect of government subsidies, and activating market-based compensation mechanisms. Moreover, this impact exhibits heterogeneity across firms with different industry attributes and greenwashing tendencies.

In summary, the contributions of this study are multifaceted. To our knowledge, this is the first micro-level study to systematically examine the relationship between CET and ESTI. We contribute to the literature in three ways: First, this study introduces LLMs into the field of energy economics and develops a semantic analysis framework for identifying enterprise ESTI based on patent texts. Using the ERNIE LLM in a supervised learning setting, we accurately classify whether a given patent embodies energy-saving effects. Second, we develop a theoretical framework that links CET to ESTI. Drawing on firm-level evidence from Chinese manufacturing, we find evidence consistent with a positive association between CET and ESTI. Our findings clarify the behavioral channel of CET and refine the literature’s focus from broad green-innovation outcomes to ESTI specifically. Finally, we elucidate the mechanisms through which CET facilitates ESTI, along three channels: heightened carbon-risk exposure, the incentive-substitution effect of government subsidies, and compensation for market failures. Under tighter emissions constraints—and given that the energy mix cannot be rapidly overhauled—these insights provide a theoretical basis for policymakers and managers to reduce reliance on fossil fuels and advance the low-carbon transition.

ESTI has emerged as a critical strategy for manufacturing firms to cope with environmental uncertainty and advance sustainable development.¹¹ In essence, ESTI represents a form of energy-biased technological progress, whereby energy use in the production process is reduced through improvements in energy efficiency and reductions in energy consumption per unit of output. Prior studies have extensively examined energy-biased technological progress. Karanfil and Yeddir-Tamsamani¹² systematically estimated translog share cost functions and applied state-space modeling techniques to investigate the conditions under which ESTI change occurs. Their results indicate that policy choices, to a certain extent, determine the direction of economic and technological change, and that ESTI change is highly sensitive to fluctuations in energy prices. Díaz and Puch¹³ not only arrived at similar conclusions but further found that high levels of investment in specific technological innovations enhance the productivity of capital inputs while simultaneously increasing energy consumption, thereby stimulating ESTI to offset the additional energy demand generated by such new investments. Based on a study of Chinese prefecture-level cities, Zhao et al.¹⁴ find that energy-biased technological progress enhances green innovation through several mechanisms, including the reallocation of innovation factors, energy transition effects, and green development effects.

On the other hand, a growing body of research has begun to examine the role of patents in characterizing ESTI, with some studies employing text analysis methods to identify the level of ESTI at the prefecture-city level. From the perspective of spatial spillovers, Xie and Li¹¹ found that climate risk exerts a positive effect on ESTI. Similarly, Xie and Chen¹⁵ demonstrated that ESTI significantly promotes urban electricity consumption, while the spatial spillover effects of urban electricity-saving technological innovation further intensify this promoting effect. Overall, the aforementioned studies still exhibit several limitations. First, existing research mainly focuses on the prefecture-city level, with insufficient attention devoted to firm-level ESTI at the micro level. Second, prior studies generally adopt electricity-saving technologies as the primary proxy for ESTI. Given the diversity of energy consumption types, relying solely on electricity-saving technologies to represent ESTI may lead to insufficient representativeness and obscure the contributions of other forms of ESTI. Consequently, existing research on ESTI measurement either remains at the relatively coarse level of energy-efficiency indicators or relies excessively on electricity-saving patents with limited representativeness, making it difficult to comprehensively capture firms’ ESTI activities. In recent years, the development of machine learning and artificial intelligence technologies has created new opportunities for constructing more refined micro-level indicators of ESTI.¹⁶

Among the various regulatory instruments, CET, as a quintessential market-based environmental regulatory tool, explicitly embeds carbon emission constraints into firms’ cost structures and revenue expectations through the mechanisms of total quantity control and quota trading. As such, it has emerged as an important policy context for examining whether regulation incentivizes innovation. On the one hand, a substantial body of research has evaluated the effects of carbon trading policies from perspectives such as carbon emission intensity,¹⁷ energy efficiency,¹⁸ and green total factor productivity.¹⁹ These studies generally find that regions and firms incorporated into the carbon trading system have achieved reductions in emission intensity per unit of output and improvements in energy efficiency, thereby providing macro- and meso-level evidence that carbon trading may drive technological progress through a mechanism of forced upgrading. On the other hand, a growing body of literature directly employs green patents, low-carbon patents, or environment-related patents as indicators²⁰ to examine the impact of carbon trading on firms’ green technological innovation. These studies generally find that firms incorporated into the carbon market exhibit significant increases in green patent output and green R&D activities. Prior studies have also explored the broader relationship between policy-driven transformation and regional economic dynamics under sustainability-oriented development frameworks.^21,22

However, existing research has, to a large extent, treated green technological innovation as a monolithic whole, with relatively little differentiation among the heterogeneous types of green technologies. In particular, insufficient attention has been paid to ESTI, which directly acts upon energy utilization efficiency. On the other hand, existing studies have inadequately characterized the decision-making mechanisms through which CET influences firms’ innovation behavior.^9,10,20,23 This may be attributable to the fact that firms do not separately disclose data related to carbon trading, whether in terms of expenditures on purchasing carbon allowances or revenues from selling saved allowances. Consequently, prior studies have found it difficult to empirically test the specific underlying mechanisms. Overall, although evidence regarding the role of CET in promoting green technological innovation is relatively abundant, a notable gap remains with respect to the more targeted question of whether and how CET can foster ESTI. This gap offers an important perspective for systematically examining the innovation effects of CET at the micro-level of firms, based on more refined measurements of ESTI.

Evolutionary game models provide a powerful framework for analyzing the simultaneous strategic choices of heterogeneous players.²⁴ We adopt this approach for two reasons. First, by capturing adaptive behavior and payoff-driven updating, it enables a clearer explanation of why CET affects ESTI specifically rather than generic green innovation. Second, evolutionary games better reflect real-world competition, in which firms’ choices respond to—and coevolve with—competitors’ decisions. Accordingly, we use an evolutionary game model to assess whether, and under what conditions, CET stimulates ESTI.

To simplify the analysis, we consider two firms producing a homogeneous good, denoted Firm 1 and Firm 2. Their unit production costs are C₁ and C₂, respectively, with C₂ = C₁+Δc. Assume a duopolistic setting with a homogeneous product. Market demand is linear: P(Q) = a-(q₁+q₂), a>0. This assumption is consistent with existing research.²⁵ The cost function is linear and thus convex, C_i(q_i) = C_iq_i, where C_i is the constant unit cost of firm i. Each firm chooses between two innovation strategies: ESTI and non-ESTI innovation. Numerous studies have shown that advanced energy-saving or green technologies typically require higher fixed investment and adoption costs.^26,27 Accordingly, we assume that the cost of energy-saving innovation is m₁ and the cost of non-energy-saving innovation is m₂, with m₁>m₂.

Meanwhile, it is assumed that the government implements a cap-and-trade CET, under which a certain quantity of carbon emission allowances is freely allocated to firms in each compliance cycle. A parameter θ∈[0,1] is introduced to characterize the degree of establishment and refinement of the carbon market: θ = 0 indicates that the carbon market has not yet been established, in which case firms’ emission-reduction behaviors cannot obtain market-based incentives; θ > 0 indicates that the carbon market has been established and is in operation, with a larger θ representing a more scientific allowance allocation mechanism, more active trading, stricter regulatory enforcement, and a more mature institutional environment. Under this setting, if a firm chooses to undertake ESTI, it can save carbon allowances and obtain emission-reduction revenues θC through the carbon market; if a firm does not implement ESTI, it will incur excess emissions, requiring it to purchase allowances from the carbon market and bear the compliance cost θC. Here, C denotes the potential revenue from saving one unit of allowance or the cost of purchasing one unit of allowance. Existing theoretical and empirical studies indicate that emissions trading, through the carbon pricing mechanism, significantly alters firms’ incentives for emission reduction and innovation,^28,29 thereby providing both empirical and theoretical support for the above assumptions.

Assume Firm 1 adopts ESTI, and Firm 2 adopts a non-ESTI strategy. This leader-follower configuration, with Firm 1 as leader and Firm 2 as follower, defines a Stackelberg game.26 The profit function can therefore be expressed as:

$${\pi }_i\left(q_i,q_j\right)=P\left(Q\right)q_i-C_i{q}_{i,}i,j=1,2$$ (Equation 1)

Taking into account the game sequence, the profit-maximization expressions of Firm 1 and Firm 2 are as follows:

$$\max {\pi }_2\left(q_1,q_2\right)=P\left(Q\right)q_2-C_2{q}_2$$ (Equation 2)

$$\max {\pi }_1\left(q_1,q_2\left(q_1\right)\right)=P\left(Q\right)q_1-C_1{q}_1$$ (Equation 3)

Given the sequence of play, the profit-maximization objectives of Firm 1 and Firm 2 are as follows:

$$q_1^s=\frac{a-2C_1+C_2}{2}$$ (Equation 4)

$$q_2^s=\frac{a-3C_2+2C_1}{4}$$ (Equation 5)

Thus, the equilibrium price can be expressed as:

$$P^s=\frac{a+2C_1+C_2}{4}$$ (Equation 6)

After incorporating carbon trading and innovation costs, the total revenue functions of both firms can be expressed as:

$${\pi }_{12}^{\left(1\right)}=\frac{{\left(a-2C_1+C_2\right)}^2}{8}-m_1+\theta C$$ (Equation 7)

$${\pi }_{12}^{\left(2\right)}=\frac{{\left(a-3C_2+2C_1\right)}^2}{16}-m_2-\theta C$$ (Equation 8)

Symmetrically, if Firm 2 becomes the leader (ESTI), the firms’ total revenue functions can be written as:

$${\pi }_{21}^{\left(1\right)}=\frac{{\left(a-3C_1+2C_2\right)}^2}{16}-m_2-\theta C$$ (Equation 9)

$${\pi }_{21}^{\left(2\right)}=\frac{{\left(a-2C_2+C_1\right)}^2}{8}-m_1+\theta C$$ (Equation 10)

If both firms opt for ESTI, a leader-follower relation may also obtain. The firms’ total revenue functions can be written as:

$${\pi }_{11}^{\left(1\right)}=\frac{{\left(a-2C_1+C_2\right)}^2}{16}+\frac{{\left(a-3C_1+2C_2\right)}^2}{32}-m_1+\theta C$$ (Equation 11)

$${\pi }_{11}^{\left(2\right)}=\frac{{\left(a-3C_2+2C_1\right)}^2}{32}+\frac{{\left(a-2C_2+C_1\right)}^2}{16}-m_1+\theta C$$ (Equation 12)

If both firms opt for traditional innovation, the interaction follows the Cournot model. The firms’ total revenue functions can be written as:

$${\pi }_{22}^{\left(1\right)}=\frac{{\left(a-2C_1+C_2\right)}^2}{9}-m_2-\theta C$$ (Equation 13)

$${\pi }_{22}^{\left(2\right)}=\frac{{\left(a+C_1-2C_2\right)}^2}{9}-m_2-\theta C$$ (Equation 14)

Assume that Firm 1 adopts ESTI with probability x and non-ESTI with probability 1-x. Similarly, Firm 2 adopts ESTI with probability y and non-ESTI with probability 1-y. The corresponding expected payoffs for Firm 1 under ESTI and under non-ESTI are as follows:

$$U_{x1}=y{\pi }_{11}^{\left(1\right)}+\left(1-y\right){\pi }_{21}^{\left(1\right)}$$ (Equation 15)

$$U_{x2}=y{\pi }_{12}^{\left(1\right)}+\left(1-y\right){\pi }_{22}^{\left(1\right)}$$ (Equation 16)

Therefore, the average expected payoff of Firm 1 is:

$$U_y=x{U}_{x1}+\left(1-x\right)U_{x2}$$ (Equation 17)

Similarly, the expected payoffs and average expected payoffs of Firm 2 can be expressed as:

$$U_{y1}=x{\pi }_{11}^{\left(2\right)}+\left(1-x\right){\pi }_{21}^{\left(2\right)}$$ (Equation 18)

$$U_{y2}=x{\pi }_{12}^{\left(2\right)}+\left(1-x\right){\pi }_{22}^{\left(2\right)}$$ (Equation 19)

$$U_x=y{U}_{x1}+\left(1-y\right)U_{x2}$$ (Equation 20)

The replicator dynamics for Firms 1 and 2 are given by:

$$F\left(y\right)=y\left(1-y\right)\left[x\left({\pi }_{11}^{\left(2\right)}-{\pi }_{12}^{\left(2\right)}\right)+\left(1-x\right)\left({\pi }_{21}^{\left(2\right)}-{\pi }_{22}^{\left(2\right)}\right)\right]$$ (Equation 21)

$$xF\left(x\right)=y\left(1-x\right)\left[y\left({\pi }_{11}^{\left(1\right)}-{\pi }_{12}^{\left(1\right)}\right)+\left(1-y\right)\left({\pi }_{21}^{\left(1\right)}-{\pi }_{22}^{\left(1\right)}\right)\right]$$ (Equation 22)

Setting F(x) = 0 and F(y) = 0, the solutions are as follows:

$$x_1=0,x_2=1,y^{\ast }=\frac{{\pi }_{12}^{\left(1\right)}-{\pi }_{22}^{\left(1\right)}}{\left({\pi }_{12}^{\left(1\right)}-{\pi }_{22}^{\left(1\right)}\right)-\left({\pi }_{11}^{\left(1\right)}-{\pi }_{21}^{\left(1\right)}\right)}$$ (Equation 23)

$$y_1=0,y_2=1,x^{\ast }=\frac{{\pi }_{21}^{\left(2\right)}-{\pi }_{22}^{\left(2\right)}}{\left({\pi }_{21}^{\left(2\right)}-{\pi }_{22}^{\left(2\right)}\right)-\left({\pi }_{11}^{\left(2\right)}-{\pi }_{12}^{\left(2\right)}\right)}$$ (Equation 24)

Equations 23 and 24 describe the population dynamics of the system. Following Friedman,³⁰ there are five local equilibria—O(0,0), A(1,0), B(0,1), C(1,1), and D(x∗,y∗)—of which only O and C correspond to stable evolutionary trajectories. In conjunction with Figure 1 and taking point D as an example, if S_ΔADBC>S_ΔADBO, the dynamic system evolves toward the stable equilibrium C; otherwise, it evolves toward O. The areas S_ΔODA, S_ΔODB, and S_ΔADBC are given by:

$$S_{\Delta ODA}=\frac{x^{\ast }}{2}=\frac{144\left[\left(m_1-m_2\right)-2\theta C\right]-2{\left(a-C_1-2\Delta c\right)}^2}{14{\left(a-C_1-2\Delta c\right)}^2-9{\left(a-C_1-3\Delta c\right)}^2}$$ (Equation 25)

$$S_{\Delta ODB}=\frac{y^{\ast }}{2}=\frac{144\left[\left(m_1-m_2\right)-2\theta C\right]-2{\left(a-C_1+\Delta c\right)}^2}{14{\left(a-C_1+\Delta c\right)}^2-9{\left(a-C_1+2\Delta c\right)}^2}$$ (Equation 26)

$$S_{\Delta ADBC}=1-S_{\Delta ODA}-S_{\Delta ODB}=1-\frac{x^{\ast }}{2}-\frac{y^{\ast }}{2}$$ (Equation 27)

Figure 1.

Evolutionary phase diagram of equilibria

The derivation of Equation 27 demonstrates that S_ΔADBC is a monotonically increasing function of θ. This indicates that, with the establishment and ongoing refinement of the CET system,^31,32 the probability that firms engage in ESTI under the impetus of the market mechanism increases. On this basis, this paper proposes the following hypothesis:

H1: CET is positively associated with enterprises’ ESTI.

As an institutional tool of environmental regulation, CET assigns tradable, commodity-like attributes to emission rights by embedding market price signals in the emissions-quota regime.³³ The CET imposes a total cap on corporate carbon emissions, transforming what was previously a cost-free emission activity into a priced and scarce resource. As a result, firms face three new categories of risk. First, compliance risk: Excess emissions require the purchase of allowances in the market, generating explicit costs. Second, price risk: Fluctuations in carbon allowance prices directly affect firms’ operating costs. Third, transition risk: Under the expectation of tightening allowance caps in the future, high-carbon assets face the risk of becoming stranded. Under such constraints, if firms maintain their existing production technologies and modes of operation, they face two options: purchasing additional carbon allowances through the CET, or reducing output.³⁴ As a result, carbon emissions become an explicit cost in the production process, significantly intensifying the legitimacy pressures confronting firms. At the same time, the uncertainty of carbon prices increases firms’ exposure to carbon risk. According to the cost-compliance theory, when faced with institutional costs or regulatory constraints, firms tend to adapt to changes in the external environment through internal adjustments.³⁵ Within the CET framework, in order to reduce their risk exposure, firms will adjust their innovation strategies, redirecting innovation toward energy conservation and thereby reducing their demand for carbon emission allowances. Building on this cost-pressure perspective, we propose the following hypothesis:

H2: The CET promotes ESTI by intensifying firms’ exposure to carbon risk.

Technological-innovation investment is associated with high costs, long time frames, and uncertain returns.³⁶ Given the coexistence of technological and environmental externalities in ESTI,³⁷ firms may lack sufficient incentives to engage in such innovation. In this context, government subsidies constitute a critical policy instrument in the early stages of ESTI. As noted by Acemoglu et al.,³⁸ R&D subsidies directly reduce firms’ innovation costs and help ESTI cross the initial threshold to breakthrough. Although many countries provide innovation subsidies for green technologies, prevailing subsidy magnitudes remain relatively low and fall short of meaningfully supporting firms’ ESTI.³⁹

Against the backdrop of tightening fiscal constraints and limited subsidy coverage,^40,41 CET activates market-based incentives. By introducing a trading mechanism, CET converts the environmental benefits of energy conservation and emission reduction into firms’ economic gains, thereby providing an endogenous impetus for ESTI.⁴² More specifically, by undertaking ESTI, firms reduce emissions, creating a surplus of emission allowances that can be sold in the carbon market to generate revenue. This revenue stream alleviates the financing pressure on ESTI and partially substitutes for government subsidies, thereby improving the economics of firms’ green transition and strengthening innovation incentives.⁴³ From the incentive-substitution perspective, we propose the following hypothesis:

H3: CET promotes ESTI by substituting for government subsidies.

Viewed through the lenses of cost pressure and incentive substitution, the foregoing mechanisms explain how CET influences firms’ ESTI choices. A further mechanism—market-based compensation—draws on institutional and market-failure theories to argue that, by operating through the external institutional environment, CET mitigates institutional impediments to ESTI and corrects allocative inefficiencies. In China, limited factor mobility, weak protection of property rights, and high institutional transaction costs have produced persistent market fragmentation. As Wu and Shi⁴⁴ note, allocative inefficiencies are widespread—especially in regions with lower levels of marketization—creating systematic barriers to cross-regional allocation of innovation resources and to technological diffusion, thereby significantly constraining the development of ESTI. Concurrently, traditional environmental regulation remains largely command-and-control, grounded in administrative directives and compulsory standards. Such inflexibility leads to inefficient allocation of regulatory effort and heightens the risk of rent-seeking.⁴⁵ The resulting institutional frictions impede factor allocation and weaken innovation incentives, relegating ESTI to the margins.

CET can mitigate these constraints through the following channels: (1) Once a firm is brought under CET, its emissions, allowance acquisition, and trading activity are monitored and disclosed within a unified institutional framework.⁴⁶ This arrangement materially narrows opportunities for rent-seeking arising from information asymmetries and administrative discretion between firms and local governments, thereby inducing a more credible green transition. (2) When property rights over emission rights are well defined, these rights are transformed from regulatory obligations into green assets with financing capacity.⁴⁷ This transformation addresses the limitations of conventional financial intermediaries in recognizing and pricing ESTI-related value, offering a sustainable market-based remedy for green firms’ financing frictions. (3) Under traditional market segmentation, firms’ ESTI efforts often encounter restricted factor mobility and weak incentives. Leveraging unified accounting standards, quota-allocation rules, and trading platforms, CET reduces interregional institutional barriers and improves the cross-regional allocation of factors—an effect that has become more salient since the launch of the national carbon market. Accordingly, the market-oriented compensation mechanism helps remedy market inefficiencies. From the market-based compensation perspective, we propose the following hypothesis:

H4: CET promotes ESTI through a market-based compensation mechanism.

In summary, this study constructs a theoretical analytical framework incorporating carbon risk exposure, the incentive substitution effect, and the market efficiency compensation effect to systematically explain the mechanisms through which CET influences ESTI, as illustrated in Figure 2.

Figure 2.

Mechanisms of the impact of CET on ESTI

Results and discussion

Validity test of ESTI indicators

A key prerequisite for the robustness of our conclusions is the validity of the ESTI measure constructed from LLMs. According to Chiou et al.,⁴⁸ green innovation can be divided into two dimensions: green process innovation and green product innovation. The essence of ESTI lies in firms achieving energy conservation and emission reduction by optimizing and improving their production processes; ESTI can therefore be expected to be associated with green process innovation. Green product innovation, by contrast, reflects firms developing new products with green attributes in response to market demand for green products, and is thus more weakly correlated with ESTI. To verify the above inference, following Xie and Zhu,⁴⁹ we employ a content analysis method based on corporate social responsibility reports to measure the levels of green process innovation (Proc) and green product innovation (Prod). We adopt the established scales proposed by Xie and Zhu⁴⁹ to measure the levels of green process innovation and green product innovation (see Table S1), and develop the scoring criteria with reference to the approach of Albertini.⁵⁰ Specifically, if a firm’s corporate social responsibility report contains relevant textual descriptions of the green process innovation and green product innovation indicators, a value of 1 is assigned; if the report provides in-depth descriptions or quantitative information on these indicators, a value of 2 is assigned; and if no description of the green process innovation indicators is provided, a value of 0 is assigned. The final value is the sum of the scores across all indicators. The validity test results for the indicator are reported in Table 1. The estimation results in columns (1)–(2) show that LNES, defined as the natural logarithm of one plus the number of energy-saving patents, is significantly and positively associated with green process innovation, whereas it exhibits no significant correlation with green product innovation, indicating that the ESTI indicator constructed in this paper is valid.

Table 1.

Validity testing of ESTI indicators

Variable	(1)	(2)
	Proc	Prod
LNES	0.3284∗∗∗	−0.0726
	(0.0829)	(0.0797)
ROA	0.4521∗∗	−0.1396
	(0.2082)	(0.1903)
Lev	−0.0885	−0.0383
	(0.1206)	(0.1093)
Size	−0.0090	−0.0082
	(0.0304)	(0.0267)
FirmAge	0.0495	0.2411
	(0.1778)	(0.1522)
Top10	−0.0321	0.1439
	(0.1602)	(0.1395)
Separate	−0.0010	0.0020
	(0.0028)	(0.0024)
Indep	0.1234	−0.2093
	(0.2964)	(0.2802)
RD	0.5183	−0.6499
	(0.5740)	(0.5481)
Constant	5.5475∗∗∗	3.9632∗∗∗
	(0.8043)	(0.6916)
Time FE	Yes	Yes
Company FE	Yes	Yes
Observations	23,740	23,740
Adj R2	0.645	0.591

(1) ∗∗∗p < 0.01, ∗∗p < 0.05, and ∗p < 0.1. (2) The figures within parentheses are the robust standard errors clustered at the enterprise level.

Baseline regression results

Table 2 reports the baseline regression results for CET and ESTI. Column (1) presents the univariate specification; Column (2) adds firm and year fixed effects; Column (3) further includes the control variables. Across specifications, the coefficient on LNES is positive and significant at the 1% level, indicating that enterprises covered by CET exhibit higher ESTI. This finding is consistent with prior studies that report a positive innovation effect of carbon trading on green innovation and low-carbon patenting.^9,10 However, unlike previous studies that focus on broad green innovation outcomes, our results specifically demonstrate that CET promotes ESTI. This finding supports H1.

Table 2.

Baseline regression results: CET and ESTI

Variable	(1)	(2)	(3)
	LNES	LNES	LNES
CET	0.7827∗∗∗	0.2469∗∗∗	0.2394∗∗∗
	(0.1236)	(0.0663)	(0.0624)
ROA	–	–	0.0245
	–	–	(0.1141)
Lev	–	–	−0.0801
	–	–	(0.0797)
Size	–	–	0.3844∗∗∗
	–	–	(0.0242)
FirmAge	–	–	0.1912
	–	–	(0.1382)
Top10	–	–	−0.0224
	–	–	(0.1189)
Separate	–	–	0.0006
	–	–	(0.0019)
Indep	–	–	−0.2201
	–	–	(0.1817)
RD	–	–	2.1410∗∗∗
	–	–	(0.3901)
Constant	1.5810∗∗∗	1.6060∗∗∗	−7.3879∗∗∗
	(0.0240)	(0.0031)	(0.6156)
Time FE	No	Yes	Yes
Company FE	No	Yes	Yes
N	24,306	24,306	24,306
Adj R2	0.014	0.783	0.795

(1) ∗∗∗p < 0.01, ∗∗p < 0.05, and ∗p < 0.1. (2) The figures within parentheses are the robust standard errors clustered at the enterprise level.

Parallel trends test

A key requirement for causal identification using DID is the parallel trends assumption. In the absence of the policy intervention, the treatment and control groups are assumed to follow the same underlying trend. Only under this assumption is it valid to use the control group as the counterfactual for the treatment group. To this end, we employ an event-study approach to estimate the year-by-year effects of firms’ inclusion in CET on ESTI. We specify the following intertemporal dynamic model:

$$LNE{S}_{i,t}={\sum }_{\tau =-6;\tau \ne -1}^8{\alpha }_{\tau }CE{T}_{i,\tau }+\gamma Control{s}_{i,t}^{\prime }+Year+ID+\epsilon$$ (Equation 28)

Here, CET_i,τ denotes the event-time indicator for enterprise i in relative year τ of inclusion in CET. The parallel-trends test is shown in Figure 3. Using the year before implementation τ = −1 as the reference period, all pre-treatment coefficients τ < 0 are statistically indistinguishable from zero, indicating no differential pre-trends between the treatment and control groups. By contrast, all post-treatment coefficients τ > 0 are significantly positive, implying higher ESTI among treated enterprises under CET. These results confirm that the parallel-trends assumption holds.

Figure 3.

Parallel trends test

points represent the estimated coefficients for each relative year of CET implementation.

Error bars indicate 95% confidence intervals.

Robustness tests

Instrumental variable approach

A potential endogeneity problem arising from reverse causality may exist between CET and ESTI; that is, firms with higher levels of energy-saving technology may be more likely to be included in the CET scheme. To address this, we employ an instrumental variable approach to mitigate the potential reverse-causality problem. We use the probability that firms in the same industry and year are included in the CET scheme as the instrumental variable for CET. The choice of this instrumental variable is based on the following two considerations: In terms of relevance, the higher the probability that firms in the same industry are included in the CET scheme, the more likely the focal firm is to be included as well, thereby satisfying the relevance condition. In terms of exogeneity, the industry-level probability does not directly affect a firm’s innovation decisions, thus satisfying the exogeneity requirement. The results in column (1) of Table 3 show that the regression coefficient of the instrumental variable is significantly positive at the 1% level. The results in column (2) indicate that the regression coefficient of CET remains significantly positive, consistent with the baseline regression results, suggesting that the core conclusions of this study continue to hold. In addition, the Kleibergen-Paap rk LM statistic is 20.978, rejecting the null hypothesis of under identification of the instrumental variable. The Wald F statistic is 29.735, exceeding the Stock-Yogo critical value of 16.380 at the 10% level, indicating that the weak instrument problem is not a concern.

Table 3.

Robustness checks

Variable	(1)	(2)	(3)	(4)	(5)	(6)	(7)
	CES	LNES	LNES_BERT	LNES	LNES	LNES	LNES
CES	–	1.1150∗∗∗	0.2335∗∗∗	0.2381∗∗∗	0.1826∗∗∗	0.2450∗∗∗	0.1701∗
	–	(0.4161)	(0.0670)	(0.0683)	(0.0607)	(0.0633)	(0.0947)
IV	0.6198∗∗∗	–	–	–	–	–	–
	(0.1137)	–	–	–	–	–	–
Constant	–	–	−7.0349∗∗∗	−7.4719∗∗∗	−7.2029∗∗∗	−7.2316∗∗∗	−7.2078∗∗∗
	–	–	(0.6099)	(0.8053)	(0.5942)	(0.6086)	(1.5869)
Control	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Time FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Company FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Industry × time FE	NO	NO	NO	NO	Yes	NO	NO
Province × time FE	NO	NO	NO	NO	NO	Yes	NO
N	24,285	24,285	24,306	24,256	24,256	24,306	5,585
Adj R2	–	–	0.777	0.795	0.813	0.801	0.890

(1) ∗∗∗p < 0.01, ∗∗p < 0.05, and ∗p < 0.1. (2) The figures within parentheses are the robust standard errors clustered at the enterprise level.

Replace the explained variable

In the baseline regressions, ESTI is measured using the ERNIE LLMs. Bidirectional Encoder Representations from Transformers (BERT) and ERNIE share the same underlying transformer-based pre-trained language model architecture, but differ in their training strategies for Chinese semantic understanding. If the ESTI indicators identified by both models support the same conclusions, the possibility of measurement error arising from the choice of a specific model can be ruled out. To assess sensitivity to the choice of LLMs, we re-measure the dependent variable with a BERT-based indicator (LNES_BERT). Column (3) of Table 3 reports estimates after replacing the dependent-variable measure. The coefficient on CET and its significance are consistent with our hypotheses, indicating that the main conclusions are not driven by the use of ERNIE.

Changing fixed effects and clustering methods

We examine the robustness of the fixed-effects and clustering specifications from three dimensions. First, the clustering level is adjusted to the industry level to address the potential correlation of residuals among firms within the same industry. Second, industry-by-year interaction fixed effects are incorporated to control for unobservable industry-level factors that vary over time, such as industry business cycles and technological cycles. Finally, province-by-year interaction fixed effects are included to account for the effects of provincial macroeconomic fluctuations and regional policy differences. The results reported in columns (4)–(6) of Table 3 show that the coefficients of CET remain significantly positive, indicating that the findings do not rely on any specific fixed-effects or clustering specification.

PSM-DID

We implement a propensity score matching–difference-in-differences (PSM-DID) approach to mitigate potential sample-selection concerns and further assess robustness. For matching, we use 1:5 nearest-neighbor and radius matching, with all control variables as covariates. In addition, following the general principles of PSM, we conduct balance tests on the selected covariates to verify the rationality of covariate selection. Before matching, the treatment and control groups exhibit significant differences at the 1% level across most covariates, indicating that the initial characteristics of the two groups are not comparable. After matching, the between-group differences in all covariates are no longer significant, suggesting that PSM effectively eliminates selection bias at the level of observable variables (see Table S2). Column (7) of Table 3 presents the estimation results based on nearest-neighbor matching, in which the coefficients are significantly positive at the 5% level, once again confirming the robustness of the conclusions.

Placebo test

To rule out the possibility that the CET-ESTI effect is a placebo arising from unobserved design factors, we implement a randomization (placebo) test. In each iteration, we randomly select a set of firms equal in size to the treated cohort, assign placebo treatment, and re-estimate the model; this procedure is repeated 500 times. Figure 4 displays the distribution of the placebo CET coefficients and their p values. The coefficients cluster around zero and are approximately normally distributed, with most p values insignificant. By contrast, the actual CET estimate lies far from this distribution, indicating that the improvement in ESTI is attributable to CET rather than random chance.

Figure 4.

Placebo test

The blue dots represent the estimated placebo-treatment coefficients and their corresponding p values obtained from 500 random assignments. The red curve represents the kernel density distribution of the placebo coefficients. The horizontal dashed line denotes p = 0.10, and the vertical dashed line denotes the actual baseline estimate of the CET coefficient (0.2394).

Evaluate the impact of omitted variables

In the baseline estimation, we include firm and year fixed effects and a rich set of firm-level controls to limit omitted-variable bias. Nevertheless, endogeneity from unobservable cannot be fully ruled out. Following Altonji et al.,⁵¹ we use selection-on-observables logic to gauge how sensitive the results are to selection on unobservable. Concretely, we compare two specifications: a parsimonious regression with a limited control set, yielding coefficient ${\stackrel{⌢}{\beta }}^R$, and a fully specified regression with all controls and fixed effects, yielding ${\stackrel{⌢}{\beta }}^F$. We construct the following index:

$$Rati{o}_{r,f}=\left|\frac{{\stackrel{⌢}{\beta }}^F}{{\stackrel{⌢}{\beta }}^R-{\stackrel{⌢}{\beta }}^F}\right|$$ (Equation 29)

As shown in Equation 30, Ratio_r,f is inversely related to ${\stackrel{⌢}{\beta }}^R-{\stackrel{⌢}{\beta }}^F$. A smaller gap between ${\stackrel{⌢}{\beta }}^R$ and ${\stackrel{⌢}{\beta }}^F$ implies greater explanatory power of the observables and a weaker influence of omitted variables on the estimates. Consequently, a larger Ratio_r,f indicates that unobservables would need to be substantially more influential to overturn the consistency of the results. Altonji et al.⁵¹ suggest that Ratio_r,f>1 indicates that the bias risk from omitted variables is relatively small. The calculation results (see Table S3) show that when the control set includes only the limited variables but incorporates firm and year fixed effects, Ratio_r,f = 6.24; when only the control variables are included without fixed effects, Ratio_r,f = 31.77. Both values far exceed the threshold suggested by Altonji et al.,⁵⁰ indicating that for the baseline conclusions to be invalidated, the correlation between unobserved factors and the carbon trading policy would need to be more than 6 times the sum of the correlations of all observable factors. Therefore, it can be reasonably concluded that the impact of omitted variables on the conclusions of this paper is negligible.

Exclusionary competitive policies

To identify the net effect of CET on ESTI, we must account for overlapping policy influences. Over 2008–2023, concurrent policies include energy saving and emission reduction demonstration city (ESER), Thousand enterprises energy conservation action (EECA), new energy demonstration city pilots (NEDC), energy use rights trading pilots (EURT), and low-carbon city pilots (LCC), among others.⁵² We construct dummy variables representing the policy shocks of the above policies and incorporate them as control variables into the baseline regression for verification (see Table S4). The results show that after excluding the competing policies, the coefficient of the core explanatory variable remains significantly positive at the 1% level, indicating that the core conclusions are not affected.

Heterogeneous treatment effects

Recent econometric literature^53,54 points out that when policy shocks occur at different points in time, the conventional two-way fixed effects DID estimator is, in fact, a weighted average of multiple 2 × 2 DID comparisons, in which already-treated earlier cohorts may be erroneously used as controls for later-treated cohorts. This can give rise to negative weighting issues and lead to biased estimates. In our sample, the carbon trading pilots were launched in multiple batches, making it necessary to examine whether such problems are present. We employ the decomposition method proposed by Goodman-Bacon⁵⁴ for verification. The results show (see Table S5) that the weights of the two types of “bad” comparisons—“Later T vs. Earlier C” and “T vs. Already Treated”—are 1.3% and 1%, respectively, exerting only a minimal influence on the overall estimation results. In contrast, the weights of “Earlier T vs. Later C” and “T vs. Never Treated” are 1.7% and 96.0%, respectively, constituting the main source of the model’s estimation results. Therefore, the estimation results of the staggered DID in the baseline regression can be regarded as reliable.

Although the Goodman-Bacon decomposition indicates that the weight on favorable comparisons is substantial—implying a very low risk of bias from heterogeneous treatment effects—we proceed prudently by estimating heterogeneity-robust DID estimators. Following the approach of Sun and Abraham,⁵³ we re-estimate the dynamic effects using an interaction-weighted estimator. The estimation results (see Figure S2) show that prior to the implementation of the emissions trading policy, the parallel trends assumption continues to hold between pilot firms and non-pilot firms. This indicates that the results obtained from the staggered DID model in examining the policy effects of CET are robust.

Mechanism analysis

Carbon risk exposure

In our theoretical analysis, we posit that CET internalizes excess carbon emissions as production costs. This increases firms’ carbon-risk exposure and induces them to implement ESTI to reduce emission-related expenses. Since firms do not separately disclose data related to carbon trading and only a small number of firms disclose carbon emissions data through their environmental, social, and governance (ESG) reports or corporate social responsibility reports, it is difficult to directly measure firms’ carbon risk exposure. To this end, we conduct textual analysis of the management discussion and analysis (MD&A) section of firms' annual reports to indirectly capture the carbon risk that firms face. Specifically, building on the study of Li et al.,⁵⁵ we construct a dictionary of corporate climate risk perception (see Table S6), and obtain firms’ carbon risk exposure by calculating the ratio of the frequency of climate risk terms to the total word count of the annual report (ClimateRisk). Column (1) of Table 4 reports the test results for the carbon risk exposure mechanism. The coefficient on CET is significantly positive at the 1% level, indicating that once firms are incorporated into the CET system, they become more sensitive to the price signals of carbon trading, thereby exhibiting a stronger energy-saving innovation effect. This finding is consistent with the view of Chen et al.,⁵⁶ namely that carbon trading can drive corporate green innovation through price signals and cost savings. We therefore conclude that firms’ carbon risk exposure is crucial to the role of carbon trading in promoting corporate ESTI. This is in line with our earlier expectation and confirms the existence of the carbon risk exposure mechanism, thus lending support to Hypothesis H2.

Table 4.

Mechanism tests

Variable	(1)	(2)	(3)
	ClimateRisk	GreenSub	ResAllocation
CET	0.0293∗∗∗	−0.0359∗	−0.0073∗∗
	(0.0092)	(0.0205)	(0.0035)
Constant	−0.2514∗∗∗	−0.1035	−0.5540∗∗∗
	(0.0811)	(0.0790)	(0.0592)
Control	Yes	Yes	Yes
Time FE	Yes	Yes	Yes
Company FE	Yes	Yes	Yes
N	23,837	24,121	21,753
Adj R2	0.826	0.412	0.259

(1) ∗∗∗p < 0.01, ∗∗p < 0.05, and ∗p < 0.1. (2) The figures within parentheses are the robust standard errors clustered at the enterprise level.

Incentive-substitution effect of government subsidies

CET provides endogenous incentives for ESTI by converting environmental benefits into economic returns. In recent years, however, subsidies for energy conservation and emissions reduction have been progressively phased out.⁵⁷ Against this backdrop, CET-generated returns can substitute for government subsidies, strengthening firms’ propensity to undertake ESTI. To verify the existence of this mechanism, we identify, from the detailed list of government subsidy items disclosed in the notes to firms’ annual reports, those subsidy items related to green initiatives—including items involving “green,” “environmental protection subsidies,” “environment,” “sustainable development,” “clean,” and “energy conservation.” We classify these as green subsidies and aggregate them to construct the corporate green subsidy indicator (GreenSub). Column (2) of Table 4 reports the test results for the incentive substitution mechanism. The coefficient on CET is significantly negative at the 10% level, indicating that CET significantly reduces firms' reliance on government subsidies. This is reflected in the fact that, even in the absence of government green subsidies, the economic incentives generated by the CET mechanism exert a significant subsidy-substitution effect on corporate ESTI. In other words, even without government financial support, firms can save allowances through ESTI and profit from carbon trading, thereby obtaining sufficient innovation returns to sustain their investment in ESTI. This result complements previous studies emphasizing the role of fiscal subsidies in promoting green innovation by showing that market-based incentives generated by carbon trading can partially substitute for government support.^40,41 This is consistent with our expected hypothesis and verifies Hypothesis H3.

Compensation for market inefficiencies

From the perspective of the external institutional environment, the market-based compensation mechanism explains how CET offsets institutional barriers that inhibit ESTI amid resource misallocation and market segmentation. Specifically, carbon trading enhances firms’ resource allocation efficiency by unifying rules, clarifying property rights, and strengthening regulatory transparency. To verify the existence of this mechanism, we draw on the approach of Richardson⁵⁸ by first estimating the firm’s reasonable level of investment for the current year, and then measuring the firm’s resource allocation efficiency by calculating the degree of overinvestment. Overinvestment leads firms to channel excessive capital and resources into projects with low rates of return, resulting in resource waste, increased opportunity costs, and higher capital costs. Column (3) of Table 4 reports the test results for the market-based compensation mechanism. The coefficient on CET is significantly negative at the 5% level, indicating that the CET scheme can promote corporate ESTI by overcoming structural inefficiencies. This finding highlights the important role of the market-oriented design of the carbon trading mechanism in remedying market failures. Carbon trading is not only an environmental governance tool but also an institutionalized compensation mechanism embedded within the market system, capable of providing sustainable institutional compensation for ESTI when institutional deficiencies exist in the market. This finding is consistent with studies highlighting the role of market-oriented reforms in improving resource allocation efficiency and stimulating innovation activities.⁵⁹ This therefore supports Hypothesis H4.

Heterogeneity analysis

Industry characteristics

The relationship between CET and ESTI may be influenced by industry characteristics. From the perspective of technological lock-in and carbon lock-in theories,^60,61 heavily polluting industries have, through long-term development, formed production systems centered on high-emission technologies. This manifests at three levels: capital lock-in (long depreciation cycles and high sunk costs of large-scale equipment), technological system lock-in (core processes are already mature and well-established, such that ESTI requires disruptive transformation), and institutional network lock-in (industrial chains and human capital have been formed around polluting technologies). This deep lock-in renders the innovation adjustment costs of heavily polluting firms far greater than what the price signal of carbon trading can cover. In short, in heavily polluting industries, polluting technologies occupy a central position within firms’ production systems, and the scope for adjusting their technological structures is limited. This results in prohibitively high transformation costs, and consequently, firms exhibit a weaker response in terms of ESTI. In contrast, the production systems of lightly polluting industries exhibit more modular characteristics, with a lower degree of path dependence, leaving broader scope for the implementation of ESTI. Based on this, we expect that the ESTI incentive effect of CET is significantly present in lightly polluting firms, while being dissipated in heavily polluting firms.

Following the secondary (two-digit) industry classification in the industry classification guidelines for listed companies revised by the CSRC in 2012, we treat the following manufacturing industries as heavy-polluting: C17 textile; C19 leather, fur, feathers (and related products), and Footwear; C22 paper and paper products; C25 petroleum processing, coking, and nuclear fuel processing; C26 manufacture of chemical raw materials and chemical products; C28 chemical fiber manufacturing; C29 rubber and plastics products; C30 non-metallic mineral products; C31 ferrous metal smelting and rolling; C32 non-ferrous metal smelting and rolling. All remaining category C manufacturing industries are classified as light-polluting. As shown in Table 5, columns (1)–(2), the CET coefficient is positive and significant at the 1% level in light-polluting industries, but insignificant in heavy-polluting industries. This pattern aligns with expectations: The impact of CET on ESTI is more pronounced among light-polluting enterprises.

Table 5.

Heterogeneity analysis results

Variable	(1)	(2)	(3)	(4)
	light pollution	Heavy pollution	True green	Greenwashing
CET	0.3181∗∗∗	0.0533	0.2506∗∗∗	−0.0133
	(0.0967)	(0.0812)	(0.0746)	(0.1036)
Constant	−6.8967∗∗∗	−8.0564∗∗∗	−7.1593∗∗∗	−4.8361∗∗
	(1.1375)	(0.7603)	(0.6586)	(1.9218)
Control	Yes	Yes	Yes	Yes
Time FE	Yes	Yes	Yes	Yes
Company FE	Yes	Yes	Yes	Yes
N	17,345	6,961	20,336	3,871
Adj R2	0.812	0.754	0.775	0.911

(1) ∗∗∗p < 0.01, ∗∗p < 0.05, and ∗p < 0.1. (2) The figures within parentheses are the robust standard errors clustered at the enterprise level.

Greenwashing

Greenwashing refers to firms’ strategic behavior of conveying exaggerated environmental commitments without undertaking substantive green actions, typically through selective disclosure and symbolic environmental communication to obtain reputational benefits.⁶² Its core characteristic lies in the divergence between environmental claims and actual environmental performance.⁶³ Compared with non-greenwashing firms, greenwashing firms tend to rely more on symbolic responses and devote fewer resources to substantive green investment and technological upgrading. Since ESTI requires sustained R&D investment and long-term technological accumulation, greenwashing firms are less likely to transform the incentives created by carbon trading into actual innovation activities. By contrast, non-greenwashing firms are more inclined to improve environmental performance through substantive technological innovation. Therefore, we expect the promoting effect of CET on ESTI to be stronger among non-greenwashing firms and weaker among greenwashing firms.

Following Hu et al.,⁶⁴ we identify corporate greenwashing by jointly examining green publicity and actual environmental performance. We gauge green publicity intensity via text analysis of the MD&A section in annual reports. If the frequency of keywords such as “green,” “environmental protection,” “low-carbon,” and “environment” exceeds the industry-year median, the firm is classified as actively green-promoting.⁶⁵ A firm is deemed to be greenwashing if, in the same year, it both qualifies as actively green-promoting and incurs environmental penalties—i.e., it projects a positive green image despite contrary performance. As reported in Table 5, Columns (3)–(4), the CET coefficient is positive and significant at the 1% level among firms without greenwashing, but statistically insignificant among firms with greenwashing. This pattern aligns with expectations: CET’s effect on ESTI is stronger when firms’ green rhetoric is matched by substantive actions.

Discussion

Overall, this study systematically investigates the relatively underexplored issue of the directional impact of CET on ESTI in Chinese manufacturing enterprises. The patent-classification procedure is illustrated with representative examples in Table 6, and descriptive statistics for the empirical variables are reported in Table 7. By integrating an evolutionary game framework with market-based environmental regulation theory and employing micro-level data from Chinese A-share listed manufacturing firms during 2008–2023, this paper uncovers the complex dynamic mechanisms through which CET influences firms’ ESTI decisions. The results show that CET significantly promotes ESTI, and the findings remain robust across multiple endogeneity and robustness tests. Mechanism analysis identifies three channels: heightened carbon-risk exposure, substitution between government subsidies and market-based incentives, and correction of market inefficiencies. Heterogeneity analysis further shows stronger effects among firms with lower pollution intensity and those without greenwashing behavior.

Table 6.

Example of energy-saving patent

Example	Patent no.	Patent abstract	Direct effect	Indirect effect
1	CN202010419752	the present invention discloses a high-performance carboxyl-terminated polyester resin for powder coatings and its synthesis method. … …, significantly reducing production costs and avoiding the problem of large energy waste caused by powder coatings synthesized by traditional polyester resins.	Yes	No
2	CN202410452067	the invention discloses a cooling circulation system and control method for an electric cotton picker. … …, by adjusting the fan’s on/off state and rotational speed according to operating conditions, regulates cooling power and ensures reliable operation of the entire machine.	No	Yes
3	CN201520538234	the utility model relates to a novel tower slant support. Including upper support ears, bent round tubes, nuts, lower support ears, and braces; … … With this structure of the utility model, the height, position, and size of the upper support ear can be adjusted, which can ensure the direct installation of the support rod of the slant support, without the need for on-site welding, saving time and labor.	No	No

Table 7.

Descriptive statistics

Variable	Variable definition	Mean	SD	Min	Max
LNES	ESTI level of enterprises measured by ERNIE model	1.617	1.401	0.000	8.271
CET	virtual variable of CET	0.047	0.211	0.000	1.000
Roa	the ratio of a firm’s total profit to total assets	0.043	0.074	−0.999	1.285
Lev	the ratio of a firm’s total liabilities to total assets	0.402	0.196	0.007	1.957
Size	the natural logarithm of a firm’s total assets	22.033	1.184	17.641	27.638
FirmAge	the natural logarithm of the number of years since the firm’s establishment	2.889	0.347	1.099	4.159
Top10	the shareholding proportion of the top ten shareholders	0.578	0.147	0.104	1.012
Separate	the control rights held by the actual controller minus the ownership rights	5.007	7.793	−55.301	60.323
Indep	the proportion of independent directors	0.375	0.055	0.000	0.800
RD	the ratio of R&D expenditure to operating revenue	0.044	0.044	0.000	0.541

In summary, this study not only deepens the understanding of the relationship between CET and ESTI, but also provides practical implications for policymakers and business managers committed to promoting sustainable development. Based on the findings of this research, the following recommendations are proposed for policymakers and decision-makers.

• (1)Promote the orderly development and expansion of a unified national carbon market. The findings of this study indicate that CET can significantly stimulate ESTI. Therefore, policymakers should gradually expand the coverage of the carbon trading system while ensuring market efficiency and regulatory capacity, incorporating a broader range of polluting industries into the trading scheme. Following the recent expansion of China’s national carbon market beyond the power generation sector to include steel, cement, and aluminum, future policy efforts should further improve sectoral coverage and gradually incorporate additional high-emission industries such as petrochemicals and chemicals. In parallel, the quota allocation mechanism should gradually transition from predominantly free allocation toward an auction-based approach. The phased expansion experience of the European Union Emissions Trading System demonstrates that gradual enlargement represents a feasible pathway for balancing market efficiency and regulatory capacity.
• (2)Optimize the government green subsidy system by establishing a complementary framework between market incentives and fiscal support. The mechanism analysis in this study demonstrates that CET partially substitutes for the incentive effects of government subsidies, implying a potential risk of functional overlap between the two policy instruments. Accordingly, it is recommended that the boundaries of fiscal subsidies be redefined based on the degree of market failure. In sectors characterized by mature technologies and sufficient market incentives, fiscal support should be gradually withdrawn in an orderly manner to avoid redundant incentives. In contrast, for frontier green technologies with high technological barriers, long R&D cycles, and substantial externalities—such as carbon capture, utilization, and storage, green hydrogen, and long-duration energy storage—carbon price signals alone are insufficient to offset investment risks, and therefore should be supplemented with differentiated fiscal support instruments.
• (3)Implement differentiated incentives and refined regulation. The heterogeneity analysis in this study reveals that the incentive effects of CET vary significantly across firms with different pollution characteristics and greenwashing tendencies, suggesting that a “one-size-fits-all” policy approach is unlikely to achieve optimal outcomes in emission reduction and innovation promotion. On the incentive side, policymakers are encouraged to deepen innovation in green financial instruments by incorporating corporate carbon performance into the core evaluation criteria for credit pricing and bond issuance, while providing differentiated financial support to growth-oriented green firms in lightly polluting industries. In addition, policymakers should explore mechanisms for directing carbon market auction revenues toward disruptive low-carbon technology R&D in heavily polluting sectors. On the regulatory side, it is recommended to establish a comprehensive regulatory closed-loop system to enhance the credibility and distinguishability of corporate green signals, thereby preventing the incentive effects of carbon trading from being diluted by strategic behaviors such as greenwashing.

Limitations of the study

This study is subject to several limitations, which also provide opportunities for future research.

First, many financial and corporate governance variables employed in the empirical analysis are generally difficult to obtain for non-listed firms. Due to data availability constraints, this study primarily focuses on listed companies. Nevertheless, this sample selection remains reasonable within the context of China’s carbon market, as enterprises covered by China’s CET pilot programs are predominantly large industrial emitters, among which listed firms occupy a dominant position in key high-emission industries. Therefore, the sample highly overlaps with the actual policy targets. As ESG disclosure regulations continue to deepen and national patent databases together with firm-level emissions data become increasingly comprehensive, future studies may utilize micro-level compliance data from CET systems to identify the innovation responses of non-listed regulated enterprises.

Second, this study employs a LLM to conduct semantic identification of corporate patent abstracts, thereby accurately extracting signals of ESTI from patent texts. Compared with traditional approaches based on International Patent Classification (IPC) codes or keyword matching, this method demonstrates significant advantages in semantic comprehension depth, identification accuracy, and domain adaptability. However, the ESTI indicator developed in this paper relies exclusively on patent text data as a single data source, capturing only explicit innovation outcomes that have been formally patented, while insufficiently identifying innovation activities that remain in the R&D stage or technologies protected as trade secrets. Future research may integrate corporate social responsibility (CSR) reports, ESG disclosure texts, and corporate recruitment data to construct a multi-source ESTI measurement framework.

Finally, the synergistic effects among different environmental regulatory instruments warrant further investigation. China’s environmental governance system is characterized by a “multi-pronged” regulatory approach, in which CET, as a market-based regulatory instrument, is often implemented in parallel with command-and-control regulations (such as environmental inspections and environmental penalties) as well as voluntary regulatory mechanisms (such as green certification and CSR disclosure). Owing to the focus and scope limitations of this study, the complementary or substitutive relationships among these three categories of regulatory instruments were not examined in depth. Future research may integrate these different forms of environmental regulation to identify the overall synergistic effects of policy combinations. Such an extension would not only deepen the theoretical understanding of China’s diversified environmental governance system, but also provide more refined empirical evidence for optimizing the coordinated design of environmental policies and avoiding the costs associated with policy overlap.

By addressing the aforementioned limitations and exploring these future research directions, scholars can further deepen the understanding of the complex relationship between CET and ESTI.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Zhang Lijie (zhanglj@xju.edu.cn).

Materials availability

The study did not generate new materials.

Data and code availability

• •All data reported in this article are from publicly available sources and are included in the article.
• •Code for this study is available at Zenodo: https://doi.org/10.5281/zenodo.20475126, or from the lead contact without restriction upon request.
• •Additional information on the data reported in this article and its analysis is available from the lead contact upon request.

Acknowledgments

This research was funded by the Western Project of National Social Science Fund of China (grant no. 24XJL012); Major Project of Social Science Fund of Xinjiang Autonomous Region (grant no. 20AZD004); Natural Science Foundation of Xinjiang Autonomous Region (grant no. 2021D01C053); Excellent Doctoral Innovation Project of Xinjiang University (grant no. XJU2024BS012). We thank all of the open-data providers and institutions for the data used in our study. We are grateful to Dr. Ryan Perry, Scientific Editor, and the two anonymous reviewers for their thoughtful evaluation and constructive feedback.

Author contributions

Y.F. conceptualization, writing—review and editing, formal analysis, and writing—original draft. J.S. conceptualization, writing—review and editing, formal analysis, and writing—original draft. K.Y. supervision, resources, validation, data curation, and writing—review and editing. L.Z. supervision, resources, validation, data curation, and writing—review and editing.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the author(s) used ChatGPT (OpenAI) for the improvement of English expression. After using this tool, the author(s) carefully reviewed and edited the content as needed and take(s) full responsibility for the content of the publication.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
ESTI	State Intellectual Property Office patent search and analysis system	https://pss-system.cponline.cnipa.gov.cn/Disclaimer
CET	CSMAR database	https://data.csmar.com/
Related control variables	CSMAR database;Chinese Research Data Services Platform	https://data.csmar.com/;https://www.cnrds.com/Home/Login
Code and processed data for this study	This Paper	Zenodo: 10.5281/zenodo.20475126
Software and algorithms
StataMP 18	Statacorp	https://www.stata.com/statamp/
Python	Python Software Foundation	https://www.python.org/
Dataset	This paper	https://doi.org/10.5281/zenodo.20475126

Method details

Data collection

This study uses A-share listed manufacturing firms in China from 2008 to 2023 as the research sample. Firms under ST or PT status were excluded, observations with missing core variables were removed, and continuous variables were winsorized at the 1st and 99th percentiles. The final sample contains 24,306 firm-year observations.

Patent information was obtained from the China National Intellectual Property Administration. Annual reports were collected from the Juchao Information Network. Financial and governance variables were obtained from the CSMAR and WIND databases. The lists of firms covered by carbon emissions trading schemes were obtained from the CSMAR database.

Measurement of energy-saving technological innovation

The abstract of a patent provides relevant information such as its application domain, functional characteristics, and technical scope. Performing semantic analysis on the abstract text can help us determine whether an innovative patent possesses energy-saving efficacy. Given the differences between Chinese and English, we adopt the ERNIE LLM, which is better suited to the Chinese-language context, to construct an identification framework for ESTI. We matched the full set of patents using the unified social credit codes of listed companies and their subsidiaries, ultimately obtaining 3.09 million patent records for Chinese manufacturing listed companies from 2008 to 2023, and then determined whether each patent possessed energy-saving efficacy.

Aligned with the connotations of ESTI, our identification proceeds along direct and indirect dimensions. In the direct case, patents whose abstracts explicitly state an energy-saving purpose (e.g., “energy conservation,” “reducing energy consumption”) are classified as ESTI. As shown in Example 1 of Table 6, the patent proposes a synthesis route for a new material that explicitly reduces energy waste, thereby qualifying as direct ESTI. Second, although the patent abstract does not explicitly mention “energy conservation,” it indicates that a particular design indirectly reduces energy use, consistent with the indirect definition of ESTI. As shown in Example 2, this patent discloses a control method for the cooling circulation system of an electric cotton picker that adjusts the radiator fan’s start/stop status and rotational speed according to the heat generated under different operating conditions, thereby modulating cooling power and ensuring stable operation of the machine. Although the abstract does not explicitly mention energy conservation, this condition-adaptive regulation of cooling power can reduce unnecessary fan power consumption (relative to constant operation), thereby indirectly achieving an energy-saving effect and meeting the criterion for indirect ESTI. Finally, because Example 3 only emphasizes improvements in structure and installation methods, it is not considered an energy-saving patent with substantial direct or indirect energy-saving effects.

As the foregoing identification procedure suggests, accurate classification of energy-saving patents requires a solid grasp of the taxonomy and linguistic patterns underlying the rules. Keyword-based text-analysis methods can reliably detect explicit, direct statements of energy saving. However, for indirect descriptions—given the complexity of patent knowledge and the heterogeneity of semantic expression—traditional approaches struggle to classify such cases effectively. Advances in large language models (LLMs) offer a viable solution to these challenges. LLMs can capture energy-saving technical attributes that are implicit in patent abstracts. In cases where indirect energy-saving effects cannot be inferred from explicit cues, LLMs substantially improve classification accuracy and applicability.

The training pipeline for LLMs comprises data annotation, model training, evaluation, and inference. First, Considering that the number of patents has increased year by year, randomly sampling a subset of patents for data annotation could result in insufficient learning of patent terminology from earlier years, thereby reducing model performance. To address this, we grouped the patents by year and extracted 1,000 patents from each year, yielding the dataset to be annotated in this study. Next, by designing “human–machine dialogue question-and-answer” prompts, we used the generative artificial intelligence models ChatGLM and ChatGPT to annotate the sampled patents, and divided them into training and test sets at a ratio of 8:2. Next, the training set was used to train the ERNIE large language model, while the test set was used to validate the model’s classification performance. Finally, the remaining patent information was fed into the model to complete the patent classification.

It should be noted that we introduce two improvements over existing research. (1) To address the inefficiency and subjectivity of manual annotation, we replace it with two mainstream generative AI models to obtain higher-quality labeled samples. We define a “patent reviewer” role (see Methods S1) and design a prompt that encodes ESTI criteria, illustrative examples, cautions, and final queries. Each patent is assessed independently by both models; concordant judgments are accepted, whereas disagreements trigger human re-verification. (2) Regarding model selection, we employ the ERNIE large language model developed by Baidu. ERNIE is a Chinese pre-trained language model based on deep neural networks. By learning from large-scale corpora, it can comprehend the semantic information embedded in patent texts, thereby determining whether they contain energy-saving features. Building upon the improvements described above, the ESTI identification framework we constructed demonstrates excellent performance. On the test set, the predicted accuracy, precision, recall, and F1-score reached 92.4%, 77.2%, 79.8%, and 78.5%, respectively. Specifically, precision reflects the proportion of patents identified by the model as ESTI that are indeed genuine ESTI, while recall reflects the proportion of actual ESTI successfully identified by the model out of all true ESTI. Given the considerable difficulty of identifying ESTI, the performance metrics presented above sufficiently demonstrate that the model exhibits strong generalization capability in classifying corporate ESTI, as well as a robust ability to identify energy-saving features in previously unseen patents. The trained ERNIE model is subsequently used to assess the energy-saving attributes of the remaining patents. These predictions are aggregated at the enterprise level to yield the number of energy-saving patents for each firm–year. The ESTI proxy, denoted LNES_i,t, is computed as the natural logarithm of one plus this count.

Measurement of carbon emissions trading

Since 2013, China has launched seven regional carbon emissions trading (CET) pilots, including those in Beijing and Shanghai. In 2021, the national carbon market was officially established, initially covering firms in the power generation sector. To accurately identify the firm-level impact of CET on innovation strategies, we obtained the list of key emitting enterprises from the CSMAR database. The treatment variable (CET) equals 1 if a firm appears on the compliance list of any carbon market in a given year, and 0 otherwise.

Control variables

To mitigate estimation bias arising from omitted variables, this paper controls for a series of variables that may influence LNES. These include return on assets (Roa), asset-liability ratio (Lev), firm size (Size), firm age (FirmAge), ownership concentration (Top10), separation of ownership and control (Separate), the proportion of independent directors (Indep), and R&D expense (RD). The descriptive statistics of the variables are presented in Table 7.

Econometric model

To test the foregoing theoretical analysis, we use Chinese listed manufacturing firms from 2008 to 2023 as our sample and estimate staggered DID model to examine the link between CET and firms’ ESTI. This approach is adopted because the implementation timing of CET pilot programs varies across provinces. The staggered DID method is capable of capturing the multi-period characteristics of the policy and identifying the average treatment effect of CET on ESTI. Compared with the conventional standard DID, it is better suited to the policy evaluation context of this paper. The baseline regression is:

$$LNE{S}_{i,t}={\beta }_0+{\beta }_{1}CE{T}_{i,t}+\gamma Control{s}_{i,t}^{\prime }+Year+ID+\epsilon$$ (Equation 30)

Specifically, LNES_i,t denotes the level of energy-saving technology innovation for firm i in year t. CETS_i,t is an indicator equal to 1 if firm i is covered by a carbon-trading scheme in year t, and 0 otherwise. $Control{s}_{i,t}^{\prime }$ is a vector of control variables. Year and ID denote year and firm fixed effects, respectively. ε is the idiosyncratic error term.

Robustness tests

Several robustness tests were conducted. First, an instrumental variable approach was used to alleviate potential endogeneity. Second, the dependent variable was replaced with a BERT-based ESTI measure. Third, alternative fixed effects and clustering methods were used. Fourth, PSM-DID was conducted to reduce sample selection bias. Fifth, a placebo test based on 500 random assignments was performed. Sixth, the Altonji ratio was used to assess omitted-variable bias. Seventh, competing policies were controlled for. Finally, Goodman-Bacon decomposition and the Sun-Abraham estimator were used to examine potential bias from heterogeneous treatment effects.

Mechanism analysis

Three mechanisms were tested. Carbon risk exposure was measured by textual analysis of the Management Discussion and Analysis section of annual reports, using a climate-risk dictionary. Green subsidies were identified from the detailed government subsidy items disclosed in annual reports. Resource allocation efficiency was measured using the degree of overinvestment based on the investment-efficiency framework.

Heterogeneity analysis

Heterogeneity analysis was conducted from two perspectives: industry pollution intensity and corporate greenwashing behavior. Manufacturing firms were divided into heavily polluting and lightly polluting groups according to industry classifications. Greenwashing behavior was identified by comparing firms’ green publicity in annual reports with their actual environmental penalty records.

Quantification and statistical analysis

All empirical analyses were conducted at the firm-year level. Continuous variables were winsorized at the 1st and 99th percentiles. Baseline regressions included firm fixed effects and year fixed effects. Standard errors were clustered at the firm level unless otherwise stated.

Statistical significance was defined at conventional levels: ∗∗∗p < 0.01, ∗∗p < 0.05, and ∗p < 0.1. The main coefficient of interest is in the staggered DID model. A positive and statistically significant indicates that carbon emissions trading promotes firms’ ESTI.

Additional resources

No additional external resources, datasets, or code repositories were used or developed for this study.