A New Perspective on Financial Risk Prediction in a Carbon-Neutral Environment: A Comprehensive Comparative Study Based on the SSA-LSTM Model

Abstract

Climate change is widely acknowledged as the paramount global challenge of the 21st century, bringing economic, social, and environmental impacts due to rising global temperatures, more frequent extreme weather events, and ecosystem disturbances. To combat this, many countries target net-zero carbon emissions by 2050, reshaping both the financial system and consumption patterns. This transition has sharpened the financial sector’s focus on climate-related risks, making the carbon footprint, environmental benefits of investments, and sustainability of financial products critical to investors’ decisions. However, conventional risk prediction methods may not fully capture these climate-associated risks in a carbon-neutral setting. Emerging from this context is the need for innovative predictive tools. Recently, Long Short-Term Memory networks (LSTM) have gained prominence for their efficacy in time-series forecasting. Singular Spectrum Analysis (SSA), effective for extracting time series patterns, combined with LSTM as SSA-LSTM, offers a potentially superior approach to financial risk prediction. Our study, focusing on a case study of the wind energy sector in China, situates itself within the growing body of research focusing on the integration of environmental sustainability and financial risk management. Leveraging the capabilities of SSA-LSTM, we aim to bridge the gap in the current literature by offering a nuanced approach to financial risk prediction in the carbon-neutral landscape. This research not only reveals the superiority of the SSA-LSTM model over traditional methods but also contributes a robust framework to the existing discourse, facilitating a more comprehensive understanding and management of financial risks in the evolving carbon-neutral global trend.

1. Introduction

Climate change is widely regarded as a significant challenge of the 21st century, with numerous scholars and organizations highlighting its far-reaching impacts on the environment, economy, and society [1,2,3]. Accompanying global temperature rise, escalating extreme weather phenomena, sea-level surge, and ecosystem degradation, profound economic, social, and environmental impacts have resonated globally. However, it is worth noting that perspectives may vary, and some individuals or groups might prioritize other pressing challenges of this century.

In this context, the financial sector is undergoing a paradigm shift, influenced by the rise of FinTech, which is reshaping financial services and fostering greater financial inclusion, especially in emerging markets. This shift, characterized by increased competition from non-traditional actors and a revolution in customer experience, presents both opportunities and challenges, including the potential to alleviate inequality and foster economic development [4].

In reaction to the climate crisis, multiple nations have adopted carbon-neutral strategies, aiming for net-zero carbon emissions by 2050. This significant transition influences production and consumption patterns, presenting both opportunities and challenges to the financial system [5]. Particularly, the financial sector, leveraging the advancements in FinTech, can play a pivotal role in steering investments towards sustainable initiatives, such as the wind energy sector in China. This sector stands as a significant player, actively contributing to the nation’s goals of achieving carbon neutrality and is characterized by its innovative approaches and substantial investments. It serves as the focal point of our analysis, offering a deeper understanding of the financial dynamics in a carbon-neutral economy.

The financial sector, traditionally perceived as both a real sector influencing economic activities and an intermediary facilitating financial transactions, has been observed to accentuate risks in the contemporary era of carbon neutrality. This observation is substantiated by a growing body of literature [5] that explores the nuanced roles and responsibilities of the financial sector in this context. Specifically, the banking sector seems to be intricately involved, with various functions such as lending and asset management being potential avenues for increased risk. This paper seeks to delve deeper into these dynamics, offering a comprehensive analysis of how and why these risks have been accentuated, drawing upon insights from both existing literature and empirical data [6]. Critical considerations in financial decision-making, particularly for investors and financial institutions, now encompass evaluating corporate carbon footprints, assessing the environmental impacts of investment activities, and scrutinizing the sustainability of financial products. This research aims to assist decision-makers such as financial analysts, investment strategists, and policymakers in navigating the complex landscape of financial risks associated with climate change. By providing a robust tool for financial risk prediction, we aim to facilitate informed decision-making that integrates environmental sustainability into financial strategies and policies [7]. However, these emerging evaluation dimensions introduce complexities to financial risk management. Scholars have ventured into diverse methodologies and models in this domain. For instance, Mello [8] executed regression studies on the debt scenarios of 44 countries from 1982–2000, suggesting the absence of a universally “optimal” strategy for debt. Additionally, Puente-Ajovin [9] employed Granger causality analysis on 29 years of fiscal data from several economic organizations, revealing significant impacts of local government debts on actual GDF growth. Notably, the International Monetary Fund, over the past 29 years, has proposed various econometric models for controlling extensive local government debt risks, with the credit risk measurement model KMV widely endorsed. In banking credit risk, Kogeda O.P. [10] ingeniously integrated artificial neural networks, validating high precision in credit risk assessment. Tan Zhongming [11] investigated the nexus between credit risk management and banking performance, discerning a close relationship. Risks in supply chain finance have also been extensively scrutinized. Devalkar [12] noted the inefficacy of vendors’ receivable strategies in managing moral risks in the supply chain, emphasizing the risk-reducing potential of core enterprises’ reverse factoring projects. Moreover, Lujie Chen [13] emphasized constructing specific risk models for supply chain finance based on quantitative indicators. In technical model research, Jairaj Gupt [14] considered the internationalization level of UK firms, studying its effects on credit risk models. Zhang L. [15] utilized SVM machine learning for supply chain finance credit risk assessment, demonstrating superior efficacy over traditional methods. Dongmei Cui [16] optimized financial credit risk assessment models with data mining and neural network techniques. Rostamzadeh R. [17] and his team introduced a novel analytical framework for supply chain financial risk management through a fuzzy hierarchy model. Ali Dehghanpour Farashah [18] emphasized holistic credit risk assessment by banks, considering all asset classes. Both local government debt management and banking credit risk management have showcased their importance within the broader milieu of carbon-neutral strategies and financial systems.

Nevertheless, conventional financial risk forecasting might be limited in a carbon-neutral setting, as it might not fully encompass climate-related risks. The imperative now is for more precise forecasting methodologies. Recently, deep learning, especially LSTM (Long Short-Term Memory networks), has shown prominence in time-series analysis [19]. When coupled with SSA (Singular Spectrum Analysis) [20], critical insights from time-series data can be extracted.

Although numerous studies have delved into the economic and social implications of climate change and the relationship between financial risk and climate change has gradually garnered attention, research specifically addressing the integration of LSTM and SSA in financial risk forecasting remains scant.

This research endeavors to intensively evaluate this novel model’s efficacy for financial risk forecasting in a carbon-neutral environment, particularly contrasting it with traditional methods. We aspire that this in-depth study offers a novel, scientific framework for financial risk forecasting amidst the current global carbon-neutral trend. Such exploration aids financial institutions in better evaluating and managing climate-related risks, fostering fintech innovation and progression. In the long run, a more robust and reliable risk forecasting framework will be essential for the stability of the global financial market and its strategies in response to climate change.

2. Materials and Methods

This study revolves around one of the greatest challenges of the 21st century: climate change and its subsequent risks and challenges to the global financial system [21]. In particular, our focus narrows down to financial risk forecasting in the context of carbon neutrality. To begin, we employ cutting-edge digital tools for literature collection and analysis pertinent to financial risk alerts in a carbon-neutral environment. Using web scraping technologies, we efficiently gather pertinent literature, followed by a rigorous text analysis to extract vital keywords, laying a solid foundation for subsequent research endeavors.

After an exhaustive literature review, our attention shifts towards the identification of an indicator system for financial credit risk warning. By setting clear principles for risk indicator selection and conducting qualitative research, we chart a clear trajectory for our study. Through literature reviews and policy text analyses, a meticulous exploration into the financial risk prediction indicators under carbon neutrality was undertaken, followed by a comprehensive assessment.

The next phase ushers in model establishment and experimentation. Specifically, we adopt the SSA-LSTM model, fusing Singular Spectrum Analysis techniques with Long Short-Term Memory network methodologies, paving the way for constructing our financial credit risk warning model. To test its utility and precision, specific industry sectors were chosen for case studies, with risk ratings categorized under a carbon-neutral backdrop. Our findings underscore the SSA-LSTM model’s efficacy and precision in forecasting financial risks.

In the end, by juxtaposing our model with alternatives, we draw a series of conclusions, summarizing our findings and presenting a novel, scientifically backed framework for financial risk forecasting in a carbon-neutral environment. Our study aspires to embolden financial institutions in their evaluation and management of climate-associated risks, thereby catalyzing innovation and progress in fintech.

2.1. Literature Retrieval and Intelligent Analysis of Financial Risk Warning under Carbon Neutrality

2.1.1. Intelligent Acquisition of Financial Risk Literature Using Web Scraping Techniques

(1)

Sources of Web Capture Data:

To ensure accuracy in building an indicator system for financial risk warning in a carbon-neutral context, this research adopted an intelligent retrieval strategy. It encompasses a spectrum of sources, ranging from academic portals, renowned news platforms, domestic financial and sector analysis websites, magazines, conferences, and more. In their study spanning from 1982–2000, Mello [8] conducted regression analyses focusing on the debt scenarios of 44 countries, critically examining the strategies adopted by these nations and suggesting the absence of a universally “optimal” strategy for managing debt. On the other hand, Puente-Ajovin [9] utilized Granger causality analysis to scrutinize 29 years of fiscal data from various economic organizations, shedding light on the significant impacts of local government debts on actual GDP growth, thereby providing a nuanced understanding of the fiscal dynamics at play. After meticulous filtering, key risk elements were ascertained. In our study, “meticulous filtering” refers to a systematic process where we employed AI-based retrieval strategies to sift through a vast array of data. This process involved utilizing various algorithms to identify and isolate the most relevant and credible data sources for our research. Specifically, we used Latent Dirichlet Allocation (LDA) as a technique to classify and group data based on underlying patterns and topics. LDA performs topic modeling, which helps in the categorization of large text corpora into distinct topics, facilitating a more focused analysis. However, we acknowledge that this method might have inherent limitations, including potential biases arising from the initial setting of topic numbers and the risk of oversimplification when complex topics are condensed into a limited number of categories. We have taken measures to mitigate these risks by cross verifying the results with other methodologies and incorporating a diverse set of data sources to ensure a comprehensive and balanced analysis.

(2)

Literature Acquisition on Financial Risks under Carbon Neutrality:

To curate a collection of articles relating to financial risks in a carbon-neutral setting, keyword searches were executed across various data sources. This ensured a high-quality corpus, incorporating terms such as ‘carbon neutrality’, ‘green finance’, ‘sustainable development’, ‘environmental risks’, ‘climate change’, and their derivatives.

2.1.2. Intelligent Keyword Analysis of Financial Risk under Carbon Neutrality Using Text Analysis

In the landscape of carbon neutrality, we embarked on an in-depth exploration of core terminologies pertaining to financial risk using text mining. Utilizing web scraping technologies, an impressive corpus of 59,382 articles was amassed. Preliminary vocabulary extraction and frequency analyses were executed on this trove. After excluding stop words, nearly 48,000 terms were curated. Given that many might be tangentially related to our core topic, we employed Latent Dirichlet Allocation (LDA) for thematic modeling, discerning truly pivotal terms [22]. Using Python’s LDA library, 2215 salient terms were identified post-analysis. After further sieving out unrelated high-frequency terms, we settled on around 480 keywords intrinsically tied to financial risks under carbon neutrality. Among these, the top 130 most representative words were chosen, grounding the creation of our financial risk warning indicator system under a carbon-neutral backdrop.

2.2. Preliminary Identification of Financial Credit Risk Warning Indicator System

2.2.1. Principles of Risk Indicator Selection

Upon systematic observation, the financial risk early warning system in a carbon-neutral environment is composed of dynamic factors including investors, financial institutions, and enterprises. In this framework, the transmission of risks between logistics, information flow, capital flow, and entities is intertwined, showcasing the complexity, diversity, and systematic nature of financial risks. We have organized the financial risks in a carbon-neutral context based on the principles of sensitivity, effectiveness, scientific rigor, holistic system consideration, and a combination of general and specific indicators, further grounded in grounded theory.

(1)

Principle of Sensitivity:

The sensitivity of early warning indicators for financial risk in a carbon-neutral environment is evident at two levels: first, the sequence of these indicators in a time series before the onset of the financial risk [23]; second, the varying strengths of warning signals, or the fluctuations in indicators before risks arise. The essence of sensitivity indicators lies in their response time and intensity.

(2)

Principle of Effectiveness:

The effectiveness of these indicators is twofold: one, the ability to accurately display and forewarn of impending financial risks; and two, significant distinctions among indicators. The selection of effective indicators should integrate theory and practice from both domestic and international financial risk early warning studies.

(3)

Scientific Principle:

The scientific principle is fundamental to indicator evaluation. Indicators must be rooted in scientific research methodologies, reflecting the unique characteristics of financial institutions and financial risks.

(4)

Holistic System Principle:

This principle requires a comprehensive examination of financial risks throughout their lifecycle. Hence, financial risks in a carbon-neutral environment should be viewed as a systems project, considering both global and specific relationships.

(5)

Combination of General and Specific Indicators:

While all financial institutions primarily act as financial intermediaries, in a carbon-neutral context, they are increasingly aligning with national climate policies, emphasizing support for low-carbon industries and businesses, and fostering green economic growth [24]. Thus, the indicator system should integrate insights from general financial risk early warning metrics, adjusting them according to the unique attributes of financial institutions in a carbon-neutral setting.

2.2.2. Objective, Sample, and Process of Qualitative Research

Given the limited research on financial risks in a carbon-neutral environment, our initial structure for the financial risk warning indicator system employs qualitative research to understand the perceptions, attitudes, and cognitions of financial institutions and credit rating subjects towards financial risk warnings. Moreover, by elucidating the essence of financial risks, we aim to offer vital references for quantitative evaluation. Our qualitative research seeks to outline the fundamental dimensions of financial risks under a carbon-neutral backdrop, investigating any potential differences between these and conventional financial risks within the Chinese context.

Grounded theory, suitable for theoretical concepts with ambiguous connotations or ongoing disputes, is our chosen method for qualitative research. Employing this approach, we delve into the structural content of financial risk warnings in a carbon-neutral setting, further harnessing the SSA-LSTM model for an in-depth comparative analysis.

2.2.3. Financial Risk Prediction Indicators in a Carbon-Neutral Environment Based on Literature Review

Deep analysis of the literature through big data methods reveals a trend towards exploring the nuances of risk management and control in a carbon-neutral environment. Financial risks originate from adverse selection, moral hazard, and rent-seeking behaviors in the credit market, all deeply ingrained in institutional structures. Wang Huang [25] analyzed the credit risk of rural supply chain financial institutions, noting the heightened market risks for small enterprises and the associated default risks, especially if poorly managed. However, under a carbon-neutral backdrop, businesses face a multifaceted risk landscape. Beyond market risks, they must navigate credit risks, operational risks, liquidity risks, and legal risks, as outlined by the BIS framework. Additionally, they must actively manage risks related to carbon emissions, introducing further complexities to their operational challenges. Predictive indicators derived from the literature review are shown in

Table 1. Indicators for Financial Risk Prediction under Carbon Neutrality (Literature Analysis).

Primary Indicators	Secondary Indicators
External Business Environment	- Relevant low-carbon policies
	- Regional GDP and carbon emission levels
	- Local government support for low-carbon development
	- Industry response to carbon neutrality
	- Competitive carbon emission level within the industry
Fundamental Enterprise Qualities	- Enterprise’s scalability towards carbon neutrality
	- Company cash flow and investment in carbon neutrality
	- Company’s debt repayment capability and carbon emission costs
	- Profitability in relation to the carbon tax burden
	- Talent pool for low-carbon technology R&D
	- Strategies and capabilities for low-carbon development
Enterprise Creditworthiness	- Credit history regarding carbon emissions
	- Tax payments, fees, and carbon emission taxation
	- Company performance in a carbon-neutral context
Enterprise Stability	- Equity structure and carbon neutrality strategies
	- Strength and implementation of low-carbon regulations
	- Top executives’ stance and risk perception towards carbon neutrality
	- Management team’s expertise and evaluation in the low-carbon sector
	- Collaborative relationships with major carbon-emitting partners
	- Carbon trade and partnerships with associated entities
Products and Market	- Company’s position within the low-carbon industry
	- Product’s carbon emission positioning
	- R&D capability for low-carbon products
	- Sales and market acceptance of low-carbon products
	- Collaborative relationships with upstream and downstream low-carbon firms

.

2.2.4. Financial Risk Prediction Indicators in a Carbon-Neutral Environment Based on Policy Texts

Principal financial management entities include banks, financial supervisory agencies, foreign exchange bureaus, and the Central Financial Institutions Council. In a carbon-neutral context, policies will significantly affect financial risks. For instance, stringent carbon emissions controls can impact businesses’ operations and financial performance, consequently influencing the risk assessment by financial entities [26].

Upon scrutinizing policy documents, we identified that credit evaluations should consider the policy landscape, capital adequacy, asset liquidity, asset quality, profitability, management quality, credit risk management, market risk management, liquidity risk management, and credit guarantees. In a carbon-neutral scenario, assessments must also incorporate a firm’s carbon emission control capacity and environmental responsibilities. The capacity for carbon emission control can be assessed by factors such as green technological innovations, investments in energy conservation and emissions reduction, and green productivity. Environmental responsibility can be gauged through a firm’s societal responsibility and environmental impact.

From an exhaustive study of the literature and policy texts, we discern that the novel perspectives for predicting financial risks in a carbon-neutral environment predominantly revolve around: green technological innovations, environmental responsibilities, policy landscape, capital adequacy, asset liquidity, asset quality, profitability, management quality, credit risk management, market risk management, liquidity risk management, and credit guarantees. These factors not only enhance the accuracy of predicting financial risks but also deepen our understanding of financial risks in a carbon-neutral context.

2.3. Comprehensive Analysis of Financial Risk Prediction Indicators in a Carbon-Neutral Environment

From the preceding literature and policy analyses, several key findings can be drawn:

(1)

Financial risk prediction in a carbon-neutral environment necessitates a focus on a company’s green transformation capability. Businesses need to prioritize both financial and environmental performance, particularly with regards to carbon emissions [27].

(2)

Technological innovation, especially in green technologies, plays a crucial role in mitigating financial risks. Investments made by companies in energy-saving and emission-reducing initiatives in a carbon-neutral context can unlock more financial and market opportunities.

(3)

The policy environment in a carbon-neutral setting profoundly impacts financial risks. Financial institutions need a comprehensive understanding of this environment to assess risks more effectively.

(4)

Beyond traditional financial risk prediction indicators such as capital adequacy, asset liquidity, and asset quality, financial entities should also evaluate novel risk indicators such as carbon emission control capabilities and environmental responsibilities.

As aforementioned, to ensure that our standards and evaluations are comprehensive and accurate, we have formulated an elaborate indicator system. This system encompasses all crucial dimensions, ensuring objective and precise evaluations, as detailed in

Table 2. Evaluation Index System of Financial Risk in a Carbon Neutral Environment.

Goal Layer	Criterion Layer	Element Layer
Evaluation Index System of Financial Risk under Carbon Neutral Environment	External Corporate Risk	Stability of Carbon Emission Industry Development
		Competitive Degree of the Carbon Emission Industry
		Government’s Preference for Low-Carbon Enterprises
		Government Support Intensity for Low-Carbon Enterprises
		Major Carbon Emission Related Party Risk
	Internal Corporate Risk	Equity Status
		Major Investment and Financing Plans, Asset Reorganization, etc.
		Enterprise Executive Risk
		Stability of Financing Operations
		Asset Quality
	Corporate Financial Status	Operational Ability
		Debt Repayment Ability
		Growth Ability
		Profitability
	Corporate Credit Record	Adverse Credit Records
		Performance of Accounts Payable
		Tax and Fee Payment Situation

.

2.4. Explanation of Indicators

(1)

External Risk of the Enterprise

In a carbon-neutral context, the external risks to enterprises mainly manifest in five areas: stability of the carbon emission sector’s development, level of competition, government preference towards low-carbon businesses, government support intensity, and risks associated with major carbon-emitting affiliates.

• Stability of Carbon Emission Industry Development

This is evaluated based on which development stage the carbon emission sector the enterprise belongs to is currently in, be it in the initiation, growth, maturity, or decline phase [28].

b.

Competition Level in the Carbon Emission Industry

The level of competition can be categorized into pure monopoly, monopolistic competition, oligopoly, and perfect competition. Businesses in the low-carbon sector, due to substantial profit margins, generally exhibit relatively lower credit risks in their bond issuances.

c.

Government Preference for Low-Carbon Enterprises

The preference of governments generally lean towards collaborations that prioritize low-carbon development to maximize environmental benefits. State-owned entities, with strong backing from the government, often have a lower risk of default.

d.

Government Support Intensity for Low-Carbon Enterprises

Amidst a carbon-neutral backdrop, government support for low-carbon businesses intensifies. Financial and policy support extended by governments enhances a business’s ability to service its debt, thereby reducing default risks.

e.

Risks Associated with Major Carbon Emission Counterparties

The primary counterparty risks arise from affiliations with high-carbon-emitting entities, such as pivotal suppliers or consumers, who might face future carbon emission caps or tax implications.

(2)

Internal Risk of the Enterprise

Under the carbon-neutral setting, the internal risks of enterprises are assessed based on their transition progress towards low-carbon, carbon emission control measures, green investment and financing plans [29], the management’s understanding and commitment towards carbon neutrality, and the quality of the enterprise’s carbon assets.

(3)

Enterprise’s Financial Condition

The financial health of a company plays a pivotal role when assessing financial risks in a carbon-neutral environment. It is primarily reflected through operational capacity, debt-servicing ability, potential for low-carbon growth, and carbon-related profitability.

(4)

Carbon Credit History of the Enterprise

In the carbon-neutral context, an enterprise’s carbon credit records, encompassing aspects such as carbon emissions, energy efficiency, and carbon trading activities, stand as crucial indicators in evaluating its financial risks.

In this section of the paper, we commence with a literature review utilizing big data, followed by keyword analysis. Applying grounded theory research methodology and based on the analyzed literature and policy documents, we identify a comprehensive indicator system for early warning of financial credit risks in a carbon-neutral context. In the selection process of these indicators, we adhere to principles of sensitivity, validity, scientific rigor, comprehensiveness, and risk assessment under carbon-neutral circumstances, ensuring the holistic and scientific nature of our choice. Results illustrate that external risks, internal risks, financial conditions, and carbon credit histories—all these four dimensions—can effectively influence the credit risks of financial institutions in a carbon-neutral environment. After iterative refinements to the indicator system, we have established the final comprehensive early warning system for financial credit risks in a carbon-neutral context.

3. Risk Prediction Model Establishment and Experiment

In this section, based on the established index system, we construct a financial credit risk early warning model utilizing the SSA-LSTM model. This model aims to identify the predominant factors affecting developmental financial credit risk. Data from specific industrial enterprises was selected and fed into the SSA-LSTM model for learning. This facilitates the evaluation and prediction of credit risk levels in diverse enterprises, determining key elements impacting developmental financial credit risk, and providing a theoretical foundation for financial risk forecasting.

3.1. Developmental Financial Credit Risk Early Warning Model Based on SSA-LSTM

3.1.1. Singular Spectrum Analysis (SSA)

Singular Spectrum Analysis (SSA) is an advanced time series decomposition technique that aims to extract meaningful information from a time series [30]. Contrary to the traditional time series decomposition methods that primarily focus on separating the time series into trend, seasonal, and noise components, SSA delves deeper, identifying and reconstructing the embedded structures within the data, ranging from oscillatory patterns to trends.

Fundamental Formulas and Components: Given a time series $X_t$, SSA primarily relies on embedding the series into a sequence of vectors, followed by Singular Value Decomposition (SVD). The decomposed series can be represented as:

$$X_t=\sum _{i=1}^r {\lambda }_i{U}_i{V}_i^T$$

(1)

where:

$X_t$ is the raw data at time t.

$r$ is the rank of the trajectory matrix constructed from the time series.

${\lambda }_i$ are the singular values.

$U_i$ and $V_i$ are the left and right singular vectors respectively.

Embedding and Trajectory Matrix: The initial phase of SSA involves embedding the time series into a multi-dimensional space, resulting in a trajectory matrix $X$. For a chosen window length $L$, the matrix is constructed by arranging lagged vectors of the series.

Singular Value Decomposition (SVD): Once the trajectory matrix is formed, SVD is applied to it:

$$X=U\mathsf{\Sigma }{V}^T$$

(2)

where $U$ and $V$ are orthogonal matrices, and $\mathsf{\Sigma }$ is a diagonal matrix with singular values.

Reconstruction: Post-SVD, grouping techniques are utilized to cluster the singular vectors, leading to the separation of different structures within the data. The series is then reconstructed based on these groupings.

Trend Component: Similar to traditional decomposition, SSA is adept at capturing long-term movements in data. This component is typically linked with the leading singular values and vectors, representing more significant structures in the data.

Oscillatory Patterns: Apart from trend and seasonality, SSA effectively captures oscillatory patterns within the series. These are intermediate components that may not have a fixed frequency, differing from regular seasonal effects.

Noise Component: Residuals or noise in the context of SSA are associated with the smaller singular values, encapsulating the random variations unexplained by the larger structures.

In conclusion, SSA offers a comprehensive framework that not only dissects time series data into familiar components such as trend and seasonality but also uncovers intricate structures embedded within. Such intricate dissection is invaluable for enhancing forecasting accuracy, refining model fitting, and offering a richer understanding of the time series dynamics in diverse scientific contexts.

3.1.2. Long Short-Term Memory Networks (LSTM)

The Long Short-Term Memory (LSTM) network is a specialized variant of the Recurrent Neural Network (RNN), specifically designed to handle sequence data with extended time intervals [31]. Traditional RNNs, when dealing with long sequences, face the challenges of vanishing or exploding gradients. LSTMs, by incorporating a unique gating mechanism, successfully circumvent these issues and have thus found extensive applications in various domains such as natural language processing, audio recognition, and time-series forecasting.

Basic Structure and Gating Mechanism

Central to the LSTM is its cell state, typically denoted as $C_t$. Apart from the conventional hidden state $h_t$, LSTM introduces this cell state, allowing the network to carry information across prolonged time spans.

LSTM Regulates the Flow of Information via Three Gates

Forget Gate: Decides what information is to be discarded or retained from the cell state.

$$f_t=\sigma \left(W_f\cdot \left[h_{t-1},x_t\right]+b_f\right)$$

(3)

Input Gate: Determines which new information will update the cell state.

$$\begin{array}{ll}{i}_t& =\sigma \left(W_i\cdot \left[h_{t-1},x_t\right]+b_i\right)\\ {\tilde{C}}_t& =\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(W_C\cdot \left[h_{t-1},x_t\right]+b_C\right)\end{array}$$

(4)

Output Gate: Decides the output value based on the cell state.

$$\begin{array}{l}{o}_t=\sigma \left(W_o\cdot \left[h_{t-1},x_t\right]+b_o\right)\\ h_t=o_t\times \mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(C_t\right)\end{array}$$

(5)

where $\sigma$ represents the sigmoid function, and $W$ and $b$ are the weights and biases, respectively, learned through training. These gate mechanisms collaboratively enable LSTMs to efficiently propagate information and gradients across long sequences.

Updating the Cell State: The updating of the cell state can be represented using the following Formula (6):

$$C_t=f_t\times C_{t-1}+i_t\times {\tilde{C}}_t$$

(6)

This ensures that the LSTM can learn to retain information over multiple time steps and decide when to incorporate or discard new information.

LSTM represents a potent variant of the RNN. It was conceptualized to address the challenges traditional RNNs face when managing long sequences. With its distinctive gate mechanisms, LSTM proficiently captures long-range dependencies, exhibiting exceptional performance across diverse tasks.

3.2. Developmental Financial Credit Risk Early Warning Model Based on SSA-LSTM

Selection and Rationale of the SSA-LSTM Model

In the assembly of an early warning index system, the selection of an appropriate model is pivotal. Given the intricacy and unstructured nature of financial risk assessment, this study opts for the SSA-LSTM model due to its proven robust applicability from prior investigations. It is imperative to understand that the essence of artificial neural networks (ANNs) lies in mimicking the behavioral traits of animal neural networks [32], constituting a mathematical model for distributed parallel information processing. The use of ANNs in financial risk research has been notably extensive. The SSA-LSTM, being a derivative model, demonstrates superior performance in certain dimensions [33]. Given the intricate, unstructured decision-making processes in financial risk assessment, the introduction of the SSA-LSTM model is akin to a timely intervention, aptly addressing such complexities. Notably, some studies even highlight its transcendence over the traditional BP neural network models when processing critical financial risk assessment techniques such as credit scoring [34].

SSA-LSTM, signifying the Singular Spectrum Analysis Long Short-Term Memory network model, is inspired by the information transmission pattern of human brain neurons. It is a multi-layer feed-forward neural network with the capability for forward propagation of input information and backward adjustment of error information [35]. Its prowess lies in learning and memorizing an extensive range of functional relationships to cater to various input-output scenarios.

More specifically, the SSA-LSTM model encompasses input, hidden, and output layers. These layers contain one or multiple neurons. During model operation, feature vectors are primarily fed into the network via the input layer. After recognition by the input layer neurons, these vectors are transferred to the hidden layer for further processing, followed by their relay to the output layer. If the output fails to meet expectations, the model initiates a backward pass and weight-adjustment mechanism, continuously self-optimizing until the output aligns with predefined conditions. This underscores the SSA-LSTMs stature as an intelligent self-adjusting model.

In conclusion, given the impressive capabilities of the SSA-LSTM model and its commendable performance in financial risk assessments, it has been chosen as the linchpin for this research. The aspiration is to construct and fine-tune a financial risk early warning index system with heightened precision.

3.3. Case Selection

3.3.1. Industry Domain Selection

In the backdrop of a push towards carbon neutrality, the significance of the clean energy sector becomes increasingly pronounced [36]. Thus, the wind energy industry has been chosen as the subject for this case study, given its potential as a renewable energy source and its growing prominence in China’s energy portfolio [37]. Contrary to earlier studies from the 1980s that suggested that wind power might be energy-negative, recent advancements in technology and infrastructure in China have potentially altered this narrative, warranting a fresh investigation [38]. Wind energy, a form of renewable energy, is a pivotal driver for achieving sustainable societal development, as substantiated by numerous studies highlighting its role in the transition towards cleaner energy sources [39]. With the intensified global regulation of traditional energy sources [40], China’s energy structure is shifting from coal reliance towards clean wind energy, a trend supported by recent policy initiatives and investments in the sector [41]. Additionally, with the maturation of wind power technology and business models in China [42], coupled with policy support [43], the wind energy sector is undergoing rapid expansion, emerging as a new growth engine for the national economy, as documented in recent economic analyses [44].

Wind power technology harnesses wind turbines to transform wind energy into electrical energy. It bifurcates into onshore and offshore wind energy. With technological advancements and cost reductions, China’s offshore wind energy is burgeoning rapidly, poised to play a crucial role in achieving carbon peak objectives. According to the National Energy Administration, China leads the globe in terms of both added and cumulative wind power capacities, with an ascending annual trend in wind power generation.

As we delve deeper into the potential of wind energy as a renewable resource and its increasingly prominent role in China’s energy portfolio, it is imperative to consider its energy efficiency. In recent years, significant advancements in wind energy technology and infrastructure improvements in China have rendered it an energy-positive sector, meaning it generates more energy than it consumes in its operations. According to recent studies and data analyses, wind energy projects in China have not only successfully amplified energy output but also demonstrated their critical role in reducing carbon emissions and fostering a transition to green energy. This not only marks a technological progression but also highlights China’s determination and efforts towards achieving carbon neutrality goals. In future research, we recommend further exploration and analysis of the efficiency and sustainability of wind energy technology to comprehend its role and impact more holistically in China’s energy structure.

During the 14th Five-Year Plan period, China projects an addition of approximately 290 gigawatts to its wind power projects, forecasting a total wind power capacity of 536 gigawatts by 2025. The brisk growth of the wind power industry chain and rapid market expansion have beckoned a surge of new entrants. The number of newly established (in operation) enterprises related to wind power has seen a significant uptick in recent years.

Thus, selecting the wind energy sector as a case not only embodies the current development trends against the backdrop of carbon neutrality but also aids in a profound understanding and addressing of potential financial risks arising from its swift growth. In the succeeding section, the SSA-LSTM model will be employed to delve deep into the financial risk forecasting issues of the wind energy sector.

3.3.2. Carbon Neutrality-Focused Financial Risk Forecast: Criteria for Sample Selection

In the context of forecasting financial risks in a carbon-neutral environment, we adopted the bi-dimensional credit rating system from the China Development Bank [45]. We set up a comprehensive set of criteria for sample selection to ensure the representativeness and diversity of the selected enterprises, based on previous studies and industry standards [46]. The criteria are as follows:

(1)

Enterprise Scale: Our focus mainly lies on large and medium-sized enterprises, as they are more resilient to systemic risks and offer comprehensive data sources. Based on the characteristics of the wind power industry, we classify companies into manufacturing and non-manufacturing sectors, each with its own scale standards [47].

(2)

Ownership Type: We consider both state-owned enterprises and private ones, as both have unique financial risk profiles [48]. State-owned enterprises, enjoying a greater degree of national policy support, remain a key client group for financial institutions. Private firms constitute 20% of our sample.

(3)

Financial Health: We evaluate a range of financial indicators of the companies, as suggested by previous financial risk assessment studies [49], including their operational capability, debt-repayment ability, growth potential, profitability, liquidity, and leverage ratios.

(4)

Regional Electricity Market: The supply and demand status of regional electricity markets are assessed, in line with the findings of [50], to determine their influence on the production and sales of the enterprises.

(5)

Power Generation Scale and Capacity Utilization: The scale of power generation and the utilization rate of production capacity are pivotal, as highlighted in [51], as they dictate the market position and cash flow magnitude, which is instrumental for our in-depth comparative study.

(6)

Operational Costs: We pay close attention to the construction and operational costs of wind energy companies, based on industry benchmarks [52], encompassing aspects such as depreciation, administrative expenses, financial costs, and repair expenses.

(7)

Management Level and Efficiency: By looking at the average number of employees per 10,000 kW, we gauge the management proficiency and efficiency of a firm, as suggested by management efficiency studies [53].

(8)

Financial Flexibility: The ability of companies to secure external financing is examined, including their capability to raise funds from shareholders, investors, lenders, and other sources.

The aforementioned eight criteria, grounded in empirical research [54], provided a robust reference for our selection of sample enterprises for an in-depth comparative study based on the SSA-LSTM model. Eventually, we chose 30 wind energy companies for our comprehensive analysis, ensuring that these firms adequately represent the industry in terms of their operational status, scale, and other relevant dimensions.

3.4. Developmental Financial Credit Risk Early Warning Model Based on SSA-LSTM

In the backdrop of carbon neutrality, models predicting financial risks require modifications and enhancements. Traditional models, such as the Credit Risk VaR model, statistical-based risk models, and data mining-based risk models [55], though effective in financial risk measurement, lack dimensions accounting for carbon emissions and environmental impact under a carbon-neutral paradigm. Light et al. [56] posited that, viewed from the lens of carbon neutrality, enterprise risks can be categorized into traditional financial risks and risks associated with carbon emissions. They further employed the Particle Swarm Optimization algorithm in tandem with Support Vector Machines to issue early warnings about the carbon-neutral risks of listed companies. Liu et al. [57] extracted carbon neutrality-related features from five aspects of borrowers, utilized random forests for feature selection, and combined with carbon emission data, developed a risk warning model suited for a carbon-neutral environment. At a provincial policy implementation level, Shandong Province, merging internet, big data, and machine learning techniques, devised an enterprise risk categorization model that factors in carbon emissions [58]. This model is capable of updating risk classification standards in real-time, enabling differentiated supervision for enterprises based on their carbon emission risk levels. Concurrently, Hebei province championed a corporate risk evaluation model in light of carbon neutrality and established an associated dynamic regulatory mechanism.

In our study, we merged indicators related to carbon neutrality with the China Development Bank’s internal risk classification system to assess enterprise risks. When carbon emissions and other risk factors of a company remain stable, the risk scoring threshold of this system sees little fluctuation. The specific divisions are as follows: scores below 26 indicate low risk; scores between 26 and 50 suggest moderate risk; scores from 51 to 75 denote medium-high risk; scores exceeding 75 are classified as high risk. Leveraging this system, we evaluated the carbon-neutral risks of 30 enterprises. Detailed results are presented in

Table 3. Risk Assessment of Enterprises Based on Credit Risk Score.

Company ID	Credit Risk Score	Risk Level
A01	40	Medium Risk
A02	15	Low Risk
A03	18	Low Risk
A04	91	High Risk
A05	68	Medium-High Risk
A06	72	Medium-High Risk
A07	79	High Risk
A08	60	Medium-High Risk
A09	45	Medium Risk
A10	64	Medium-High Risk
A11	63	Medium-High Risk
A12	53	Medium-High Risk
A13	20	Low Risk
A14	45	Medium Risk
A15	55	Medium-High Risk
A16	93	High Risk
A17	86	High Risk
A18	28	Medium Risk
A19	36	Medium Risk
A20	36	Medium Risk
A21	50	Medium Risk
A22	41	Medium Risk
A23	60	Medium-High Risk
A24	82	High Risk
A25	63	Medium-High Risk
A26	13	Low Risk
A27	46	Medium Risk
A28	94	High Risk
A29	58	Medium-High Risk
A30	30	Medium Risk

.

3.5. Developmental Financial Credit Risk Early Warning Model Based on SSA-LSTM

In our study, we initially segmented the dataset to ensure both training and testing data were available. Specifically, samples labeled 1–25 were designated for training, while samples 26–30 were set aside for testing. To effectively train the SSA-LSTM model and ensure its predictive accuracy, we undertook the following hyperparameter settings and adjustments:

Learning Rate: A learning rate of 0.1 was set to control the speed of weight adjustments during the model’s learning phase.

Regularization: Within the SSA-LSTM model, the Bayesian regularization algorithm was utilized to counteract overfitting. This algorithm operates by adding a regularization term to the original cost function, usually related to the squared sum of network weights. This adjustment transforms the cost function from a simple mean squared error (MSE) to:

$$\mathrm{Cost} \mathrm{Function}=MSE+\alpha (\mathrm{Squared} \mathrm{sum} \mathrm{of} \mathrm{network} \mathrm{weights})+\beta$$

(7)

The optimal ratio for $\alpha$ and $\beta$ can be denoted when the cost function is minimized as:

$$\mathrm{Optimal} \mathrm{Ratio}=\frac{Ns}{n}$$

(8)

Here, $Ns$ stands for the number of training samples (25 in this study), and n signifies the total neurons within the network.

Neural Network Architecture: The SSA-LSTM model is composed of three layers: input, hidden, and output.

Neuron Count per Layer: The output layer encompasses 17 feature vectors, corresponding to all the indicators. The outcome of the output layer, the financial risk level, consists of just one neuron. The structure and neuron count of the hidden layer directly influence the model’s training effectiveness and speed. After thorough evaluation, the neuron count for the hidden layer was determined to be:

$$\mathrm{Neuron} \mathrm{count} \mathrm{in} \mathrm{hidden} \mathrm{layer}=x+y\times a$$

(9)

Here, $x$ represents the neuron count in the input layer, $y$ signifies the output layer’s neuron count, and $a$ is a number between 1 and 10. Following this formula, the neuron count for the hidden layer was set at 15.

Iterations: For training efficiency and model stability, we opted for 1000 iterations.

Activation Function: Given its characteristics and the SSA-LSTM model’s features, the sigmoid function was selected as the activation function.

Training Function: The train/m function was employed for model training.

Training Objective: To ensure the model not only converged during the learning phase but also optimized predictive accuracy, the training goal was set to less than 0.01.

Through the described tuning process, our objective was to ensure the SSA-LSTM model achieved optimal predictive performance while ensuring model stability and generalizability.

Detailed experimental results are illustrated next. Figure 1 shows the change of the loss function during model training, and Figure 2 shows the comparison between the predicted value and the true value.

In our analysis, the X-axis labeled “time” represents discrete time intervals during which the data were collected and analyzed. Each point on this axis corresponds to a specific period, which, for the purpose of this study, we have denoted in months. This chronological representation allows us to meticulously track and analyze the trends and patterns in the financial risk variables over a sustained period. This time-series data are crucial in training our deep learning model, enabling it to identify potential patterns and make more accurate predictions for future intervals. We will ensure to annotate the specific time periods on the X-axis in the revised version of the figures to provide a clearer understanding of the temporal scope of our analysis.

The Y-axis in Figure 2 and Figure 3 denotes the values of the ‘corporate loans past due over 90 days’ metric, a significant indicator in financial risk analysis. This metric is represented as a percentage, with the scale ranging from 0 to 120 to encompass the entire range of values observed in our dataset. We acknowledge that the representation of negative values in the graph needs further clarification. These values can occur due to the normalization techniques applied during the data preprocessing phase, which is a common practice in deep learning computations to enhance model stability and performance. We will revisit our data processing steps to provide a more detailed explanation of this phenomenon in the figure captions, ensuring a comprehensive understanding of the depicted values.

The primary financial risk variable under scrutiny in our study is the percentage of ‘corporate loans past due over 90 days’. This metric serves as a reliable indicator of potential financial instability within a corporate entity, thereby being a focal point in our predictive analysis. Our deep learning model, specifically the SSA-LSTM model, is designed to forecast the future values of this variable, offering a powerful tool for financial analysts and stakeholders to anticipate potential financial downturns with greater accuracy. This, in turn, facilitates more informed decision-making in financial management and policy formulation. We will elucidate this aspect further in the manuscript to provide readers with a clear understanding of the financial risk variable being analyzed and its significance in the context of our study.

3.6. Results Analysis

In an effort to further ground our analysis in real-world applications, we would like to clarify that the ‘predicted’ and ‘true’ values represented in Figure 2 and Figure 3 are indicative of specific financial metrics critical to the industry, particularly focusing on the metric of ‘corporate loans past due over 90 days’. This metric is a significant indicator in financial risk analysis, often serving as a precursor to identifying potential financial instability within a corporate entity. By focusing on this metric, our model aims to provide financial analysts and stakeholders with a tool that can potentially forecast financial downturns with greater accuracy, thereby allowing for more informed decision-making in financial management and policy formulation. We believe that this clarification not only enhances the comprehensibility of our research but also significantly elevates its applicability and relevance in the current financial landscape.

Following numerous iterations of training the SSA-LSTM model, we conducted a meticulous examination of the training and validation losses, arriving at the ensuing pivotal conclusions:

• Evolution of Training and Validation Losses: Data indicated that with the rise in iteration count, both training and validation losses manifested a general decline. This underscores the model’s gradual optimization of parameters during the learning process, enhancing the fit to the training data. However, the subtle elevation in validation loss observed during certain epochs suggests potential adaptability issues or signs of overfitting for the model under specific conditions.
• Model Evaluation Metrics: To offer a more nuanced appraisal of the model’s performance, we inspected the following cardinal metrics:

• RMSE (Root Mean Square Error): A standard metric in regression analyses, RMSE gauges the size of the predictive error, signifying the discrepancy between observed and predicted values.

$$\mathrm{RMSE}=\sqrt{\frac{1}{n}\sum _{i=1}^n {\left(y_i-{\hat{y}}_i\right)}^2}$$

(10)

where $y_i$ represents the observed value, ${\hat{y}}_i$ is the model’s predicted value, and $n$ denotes the count of observed values. The RMSE value for this study stood at 11.4997484297466.

• MAE (Mean Absolute Error): This signifies the average absolute difference between the predicted and actual values.

$$\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{n}\sum _{i=1}^n \left|y_i-{\hat{y}}_i\right|$$

(11)

where, again, $y_i$ is the observed value, ${\hat{y}}_i$ is the model’s predicted value, and $n$ represents the count of observed values. This study’s MAE value was measured at 8.292317753159216.

• R² (Coefficient of Determination): This metric expresses the ratio of variance predicted by the model to the total variance.

$$R^2=1-\frac{S{S}_{\mathrm{res}}}{S{S}_{\mathrm{tot}}}$$

(12)

where $S{S}_{\mathrm{res}}$ stands for the sum of squared residuals and $S{S}_{\mathrm{tot}}$ is the total sum of squares. An $R^2$ value closer to 1 implies the model’s superior explanatory capacity. The $R^2$ value for this research was 0.8927849244458038.

In conclusion, the SSA-LSTM model showcased commendable performance in this study, especially bearing an impressively high explanatory capacity on the $R^2$ metric. These evaluative metrics provide a comprehensive and profound perspective, aiding in understanding and gauging the model’s predictive efficiency, as well as its invaluable applicability in assessing financial risks in a carbon-neutral milieu.

3.7. Developmental Financial Credit Risk Early Warning Model Based on SSA-LSTM

While deeply investigating financial risk predictions, sole reliance on a singular model is seldom sufficient. Opting for diverse models for experimentation and subsequent comparison aids us in discerning the data characteristics from various angles, pinpointing the most appropriate methodology for the issue at hand. By juxtaposing different models, we can gauge the robustness, predictive accuracy, and computational efficacy of each, furnishing formidable support for their real-world application.

Here is a succinct introduction to the models we have opted for:

(1)

SLP (Single Layer Perceptron): This embodies the simplest form of a feed-forward neural network. The SLP comprises just an input layer and an output layer, formed by a single neuron or multiple neurons. Although SLPs can handle linearly separable data, their efficacy tends to wane when grappling with intricate, nonlinear data challenges.

(2)

MLP (Multi-Layer Perceptron): In contrast to the SLP, MLP incorporates one or more hidden layers, enabling it to learn and depict more complex functions. Endowed with heightened expressive capability, especially for nonlinear functions, it can approximate any continuous function given appropriate training.

(3)

LSTM (Long Short-Term Memory): This represents a specialized type of Recurrent Neural Network (RNN) and is particularly adept at handling and predicting salient events in time series with extended intervals and lags. The unique gating structure of LSTM effectively counteracts the long-term dependency and gradient vanishing issues encountered by traditional RNNs.

Pitting these models against each other in our experiments elucidates how differing model structures and complexities impact predictive prowess. For instance, while the SLP may be apt for straightforward datasets, it might falter amidst more complex ones. Conversely, due to their intricate structures, MLP and LSTM generally excel at more intricate tasks. Such comparisons empower us with a more holistic grasp of the strengths and weaknesses of each model, along with their applicability in distinct scenarios.

The results from the model comparison experiments are elucidated in

Table 4. Model evaluation index.

Models	RMSE	MAE	R2
SLP	11.6252383673	8.7862753011	0.8904322081
MLP	11.0747477848	8.2172155111	0.9005632628
LSTM	10.9989345739	7.7295231344	0.9019200009
SSA-LSTM	10.6999910537	7.5702516568	0.9071790507

. Figure 3 shows how the predicted values of the different models compare to the true values. In Table 4 and Figure 3, the ‘estimated’ values represent the predictions generated by the models based on the input data, while the ‘true’ values signify the actual observed values in our dataset, serving as a benchmark to evaluate the predictive performance of our models. The X-axis delineates the different instances or time points in the dataset, and the Y-axis denotes the values of the financial risk indicators that are being predicted. To foster a clearer understanding, we have included a legend in the chart that distinctly indicates what each axis represents and explicates the significance of the ‘true’ and ‘estimated’ values. Furthermore, we have incorporated a section in the manuscript elucidating the methodology employed to derive these values and the criteria used for evaluating the models through metrics such as RMSE, MAE, and R².

Upon examining the empirical outcomes of the aforestated models concerning the financial risk prediction task, the following insights can be garnered:

(1)

RMSE (Root Mean Square Error): All models exhibit RMSE values within a comparable range. This indicates that the mean deviation between the models’ predictions and the actual values lies within similar numeric confines. Given the data, the SSA-LSTM model boasts the lowest RMSE value, suggesting superior predictive accuracy relative to its counterparts. Conversely, the SLP model, with the highest RMSE, trails slightly behind other models in terms of accuracy.

(2)

MAE (Mean Absolute Error): The SSA-LSTM model manifests the lowest MAE, signifying the least average deviation in its predictions. On the contrary, the elevated MAE of the SLP model reaffirms its subpar performance in this task compared to the more intricate models.

(3)

R² (Coefficient of Determination): R² serves as a pivotal metric in evaluating model predictions, ranging between 0 and 1; the closer its value to 1, the better the model’s predictive power. As per the provided data, the SSA-LSTM model’s R² peaks at 0.9071, denoting top-tier predictive prowess. While the SLP model’s R² stands at 0.8904, marking the lowest among the four models, it remains within an acceptable margin.

Synthesizing the above results, we can infer that, specific to this study’s financial risk prediction task, the SSA-LSTM model outperforms in terms of predictive performance. This superiority possibly stems from its amalgamation of LSTMs sequential learning capabilities and SSAs time-series decomposition strengths, thus exhibiting heightened prowess in capturing intricate patterns within data. The simpler SLP model, notwithstanding its structural simplicity and lucidity, seems somewhat incapable of handling such intricate tasks. The performances of MLP and LSTM lie between SLP and SSA-LSTM. While both qualify as sophisticated models, they marginally trail SSA-LSTM in this particular endeavor.

In summation, selecting the right model is pivotal to the success of predictive tasks. Through a comparative analysis of diverse models, we can discern the strengths and weaknesses of each, furnishing invaluable insights for future research and applications.

4. Conclusions

This paper presents conclusions and recommendations derived from our study on financial risk early warning models in the context of carbon neutrality, paving the way for further research in this critical area.

Some of our key findings, which we believe can spur further research in this domain, include:

(1)

Inextricable Link between Financial Risk and Carbon Neutrality: The current global inclination and policies towards carbon neutrality underscore the urgency of accurate financial risk forecasting. Corporations’ operational environments, technological innovations, and climate change adaptation strategies are direct determinants of their financial health and the associated credit risks for financial institutions.

(2)

Influence of Internal and External Environments on Enterprises: Both internal organizational structures and external factors, such as carbon emissions control, technological innovation, and policy support, play pivotal roles in sounding early warnings of financial risks. Notably, within a carbon-neutral setting, an entity’s internal scenarios, strategic planning, financial health, and credit histories become essential reference points for financial institutions’ risk assessments.

(3)

Advantages and Application of the SSA-LSTM Model: The SSA-LSTM model, blending Singular Spectrum Analysis with Long Short-Term Memory networks, boasts exemplary precision and adaptability when forecasting financial risks within a carbon neutrality backdrop. Compared to alternative financial risk prediction models, the SSA-LSTM better captures intricate temporal sequence traits and portrays genuine risk conditions.

(4)

Role of Reputation Mechanisms in Risk Mitigation: Enterprises’ carbon-neutral performances impact not only their operational results but also their long-term standing in the financial markets. Consequently, for financial entities, an enterprise’s carbon-neutral endeavors and reputation are fast emerging as essential criteria for credit risk assessments.

Based on our study, we have the following recommendations:

(1)

Promote Intersectoral Collaboration: Financial institutions, governments, regulatory bodies, and businesses must bolster their collaborations. Together, they should confront the financial risks that emerge from the carbon-neutral backdrop and jointly devise more scientific and pragmatic risk mitigation strategies.

(2)

Deepen Research and Practical Application: Although the SSA-LSTM model and the reputation mechanism exhibit supremacy in both theory and empirical studies, their real-world applications necessitate ongoing optimizations and fine-tuning. It is hoped that future research teams and practitioners will refine these tools further, offering financial institutions more precise and functional risk management solutions.

(3)

Enhance Social and Environmental Consciousness: Amidst the larger narrative of carbon neutrality, both businesses and financial institutions need to heighten their cognizance of environmental and societal responsibilities, incorporating these facets into the heart of their operations and risk management. Future research could delve into how societal perceptions of carbon neutrality influence financial markets and how these perceptions can be quantified and incorporated into risk models, possibly exploring new methodologies and frameworks for this purpose.

In essence, this research has proffered a comprehensive theoretical and empirical analysis concerning financial risks in a carbon-neutral scenario. Using the SSA-LSTM model, we elucidated its unique advantages in predicting intricate financial risks and accentuated the paramount role of the reputation mechanism in risk assessments. Financial risk management within a carbon-neutral setting is an emerging and deepening research domain, entailing synergistic interplay among financial institutions, governmental regulatory bodies, enterprises, and other pertinent entities. The results from this study are not only instructive for the academic community but also equip financial practitioners with viable risk assessment and management strategies. We encourage future researchers to build upon our findings, possibly exploring comparative studies with other predictive models and investigating the evolving dynamics of policy, technology, and finance in a carbon-neutral world.

Specifically, future research could delve deeper into the following areas:

(1)

Developing Advanced Algorithms: Investigating the potential of integrating newer machine learning and artificial intelligence algorithms to enhance the predictive accuracy of financial risk models.

(2)

Sector-Specific Analysis: Conducting sector-specific studies to understand the varying impacts of carbon neutrality policies on different industries and developing tailored risk assessment strategies for each sector.

(3)

Policy Impact Analysis: Analyzing the impact of emerging global and regional policies on carbon neutrality and how they influence financial risks and corporate strategies.

(4)

Technological Innovations and Financial Risks: Exploring the role of technological innovations in shaping financial risks, particularly focusing on the advancements in renewable energy sources and their implications on financial markets.

(5)

Case Studies and Comparative Analysis: Conducting case studies on corporations with different carbon-neutral strategies and performing comparative analyses to identify the most effective strategies for mitigating financial risks.

It is imperative, however, to recognize that as global carbon-neutral strategies evolve and financial technologies continually innovate, risk prediction models and strategies must undergo persistent revisions and refinements. We envision a vibrant future research landscape where investigations delve deeper into the intricate interplay between financial risks and climate change, exploring how the latest technological advancements can be more effectively incorporated into risk identification, evaluation, and management, ultimately contributing significantly to financial stability and sustainable growth.