Analyst Logical Inconsistency and Stock Price Crash Risk: Evidence from Large Language Models

Abstract

This paper examines whether and how analyst logical inconsistency affects the stock price crash risk in China’s A-share market. The sample includes 3784 listed non-financial companies. The final sample comprises 12,460 firm-year observations from 2016 to 2023. Specifically, we employ the open source Qwen1.5-14B-Chat model, an instruction-tuned generative large language model, to measure analyst logical inconsistency, classify the sentiment expressed in analyst reports, and determine the differences between earnings forecasts and textual tone. Using a panel fixed-effect model for basic regression, and applying two-stage least squares and propensity score matching to deal with endogenous problems, we find that analyst logical inconsistency significantly increases the stock price crash risk. The results remain robust to alternative variable definitions, additional control variables, alternative sample periods, and more stringent fixed-effects specifications. The mechanism test shows that the analyst logical inconsistency increases the stock price crash risk through three channels: increased financial risk, reduced investment efficiency, and degraded information disclosure quality. Heterogeneity analysis also shows that this positive impact is strongest in companies with high media coverage, good corporate reputation and low ESG performance. Our research results are helpful to the study of information intermediaries and stock price crash risk by introducing measures for the quality of analyst reports based on large language models, and provide useful suggestions for regulators and investors in emerging markets.

1. Introduction

After the international financial crisis, the negative fluctuations in stock prices and the resulting economic consequences have gradually become an important topic in capital market research. The distribution of stock market returns exhibits significant asymmetry. Specifically, the impact of negative information on stock prices is significantly greater than the upward effect brought by positive information of the same intensity (J. Chen et al., 2001; Karasan et al., 2025). This asymmetry further leads to a significant negative skewness in the distribution of stock returns. In this context, stock price crash risk, as an extreme tail risk, is typically manifested as a sudden and unexpected sharp decline in stock prices over a short period (Hutton et al., 2009). A primary explanation for stock price crash risk is firms’ withholding of negative information (Zhang & Wang, 2025). Using information asymmetry theory, Jin and Myers (2006) put forward the idea of “information opacity” and agency conflict. They argue that executives often conceal negative information at work for reasons such as professional reputation, salary or contracts. Investors may view a specific stock price crash as a separate event and consider that there is uncertainty regarding fundamental risks and expected cash flows (Roy et al., 2025). This behavior exacerbates information asymmetry between firms and outside investors.

Logical consistency is an important dimension of analyst report quality and is therefore of concern to investors and financial institutions (Bradshaw, 2012). A firm is usually tracked by multiple analysts, and the degree of logical consistency may vary across analyst reports to some extent. The main reason for logical inconsistency is that analysts face a trade-off between optimism and accuracy. Optimistic analyst reports can increase analysts’ income by stimulating stock trading, while accurate analyst reports can enhance analysts’ professional reputations (Jackson, 2005; Gu et al., 2013). In an ideal, well-functioning market, analysts act as rational information intermediaries by identifying risks and communicating relevant signals through coherent reasoning and forward-looking judgments. In addition, there is also a widespread logical inconsistency between the key audit matters and the audit opinions, which may undermine the credibility and readability of the report (Duan & Huang, 2025). The occurrence of logical inconsistency in analyst reports may create information confusion.

Based on this, this paper employs an instruction-fine-tuned large language model to accurately identify the tone in analyst reports, to construct an indicator of analyst logical inconsistency, and to systematically examine its impact on stock price crash risk. At the theoretical level, this paper builds a transmission mechanism framework from three dimensions: financial risk, investment efficiency, and information disclosure. To empirically test the influence of analyst logical inconsistency on stock price crash risk and its underlying mechanisms, we employ a two-way fixed-effects panel model to estimate the baseline relationship and conduct subsample analyses to examine heterogeneity across media coverage, corporate reputation, and ESG performance. The main contributions are as follows: First, we innovatively employ a large language model fine-tuning method to identify analyst sentiment, and construct a deviation index system from three aspects: valuation conclusions, textual features, and logical consistency. Through multi-level quantitative indicators, we conduct an in-depth examination of analyst logical inconsistency. Second, we incorporate analyst logical inconsistency into the analytical framework of stock price crash risk determinants, finding that it has a significant promoting effect. This broadens the scope of research into the factors that lead to stock market crashes and offers concrete evidence on how information management can be a key player in mitigating the risk of such crashes. Third, we reveal the logical chain through which analyst logical inconsistency elevates stock price crash risk by increasing corporate financial risk, reducing corporate investment efficiency, and lowering corporate information disclosure quality, thereby deepening the understanding of corporate governance effectiveness and its realization mechanisms. The above work not only enriches the cross-disciplinary research on the economic consequences of analyst reports and stock market risks, but also provides useful policy implications for improving the governance of analyst report standards and promoting the healthy and stable development of the stock market. The remainder of this paper is organized as follows. Section 2 reviews the related literature and develops the research hypotheses. Section 3 describes the research design, including the sample selection, variable definitions, and empirical models. Section 4 presents the empirical results, including the baseline regressions, endogeneity tests, robustness checks, mechanism analyses, and heterogeneity analyses. Section 5 concludes the paper and discusses the theoretical and practical implications, limitations, and directions for future research.

2. Literature Review

2.1. Direct Impact of Analyst Logical Inconsistency on Stock Price Crash Risk

L. J. Ma et al. (2022) defined the phenomena of “optimistic earnings forecast + pessimistic textual interpretation” and “pessimistic earnings forecast + optimistic textual interpretation” in analyst reports as analyst logical inconsistency. From a functional perspective, analysts serve as information collectors and transmitters in capital markets and help alleviate information asymmetry between listed firms and investors (T. Chen et al., 2015). However, in actual capital markets, influenced by cognitive biases, information processing capability constraints, and incentive mechanisms, analyst reports generally contain certain forecast deviations and textual sentiment deviations. Such deviations may stem from judgment errors under incomplete information, and may also reflect strategic expressive behavior (Nardi et al., 2022). Existing studies generally measure textual sentiment using dictionary-based methods or conventional classification models such as BERT. Dictionary-based methods mainly identify sentiment by counting positive and negative words. Although these methods are transparent and relatively easy to implement, they may overlook contextual meaning, negation, conditional statements, intensifiers, and the distinction between descriptive and evaluative language. Conventional BERT-based models can capture contextual information more effectively than word-counting approaches, but their performance is still dependent on the specific training data, classification labels, and task structure.

Large language models can effectively address judgment errors. They possess more objective and powerful information processing capabilities. Therefore, this paper attempts to utilize large language models to construct an indicator for analyst logical inconsistency, thereby enhancing the accuracy of the indicator’s portrayal (Yang et al., 2025). From the perspective of the information formation mechanism, analysts integrate public information and potential non-public information to form a comprehensive judgment of corporate value. In an ideal effective information environment, analyst forecasts and textual interpretations in the market should gradually converge through sufficient information competition and repeated revisions, reflecting a relatively consistent value judgment. However, significant disagreements persist among analysts in the real market. Such disagreements not only reflect differences in information acquisition and interpretation capabilities but also reflect, to a certain extent, the uncertainty and complexity of the information environment (C. Kim et al., 2017). The negative effects of analyst forecast deviations on stock price crash risk mainly manifest in two ways: First, analysts, driven by self-interest or private information, issue reports inconsistent with those of other analysts, thereby exacerbating information asymmetry in the stock market and reducing investors’ understanding of the company’s actual situation, thus leading to higher stock price crash risk. The LLM based measure is particularly relevant in this context because it captures not only whether positive or negative words appear in a report but also whether the overall narrative is consistent with the direction of the earnings forecast. This makes it possible to identify information conflicts that may otherwise be misclassified as ordinary analyst disagreement or overlooked entirely. Second, analyst forecast deviations indicate that analysts hold an uncertain attitude towards the company’s future development. Analysts transmit this uncertainty to investors, thereby bringing greater uncertainty to stock prices and leading to an increase in future stock price crash risk (He et al., 2019).

From the perspective of classical crash risk formation mechanisms, if negative information in the market fails to be timely and adequately reflected in stock prices through effective trading mechanisms, stock prices may appear overvalued for a period; when negative information continues to accumulate and exceeds a certain threshold, this information tends to be reflected in market prices in a concentrated manner, thereby significantly increasing stock price crash risk (Hong & Stein, 2003). From the perspective of information asymmetry theory, the logical inconsistency in analyst reports will further aggravate the information differences and cognitive differences among market participants. Institutional investors usually have better information integration and analysis capabilities. They can systematically evaluate the historical accuracy of analysts’ forecasts, their rating change behavior and profit forecast results in a multi-source framework, and filter and correct unreliable information. At the same time, they can reconcile conflicting information between different information dimensions in the report, so that they can make relatively rational investment decisions. On the other hand, due to the limitations of their information processing ability and professional knowledge, it is often difficult for retail investors to identify the internal contradictions between different information dimensions in the analysis report. They are more likely to rely on clear and easy-to-observe information signals, such as the magnitude of rating changes or the intensity of emotional expression, and are less likely to deeply analyze basic information, such as earnings forecasts. This asymmetry in the information dependence structure makes it easier for retail investors to make irrational responses to logically inconsistent reports and take more extreme trading behaviors.

In capital markets, due to investor herding, this behavioral difference becomes greater. When investors develop different interpretations of the same analyst report, the level of heterogeneous beliefs in the market rises and trading disagreements intensify, thereby increasing short-term stock price volatility and exacerbating deviations from fundamentals. From the perspective of information accumulation and release mechanisms, logical inconsistency in analyst reports may, to a certain extent, reflect analysts’ selective expression and structural presentation of negative information during information processing. That is, through the reorganization and differential emphasis of information, some negative information fails to be fully disclosed. This mode of information expression may lead to the gradual accumulation of negative information in the market in a dispersed and implicit manner, rather than being fully priced at an early stage. Under persistent information asymmetry and reinforced heterogeneous interpretations, the real operational risks and negative information of firms are difficult to be timely and adequately reflected in stock prices, thereby forming a phase of price overvaluation. When accumulated negative information is centrally identified by the market or forcibly released during external shocks or information re-disclosure processes, it often triggers relatively severe price adjustments, thereby significantly increasing stock price crash risk (L. J. Ma et al., 2022). Based on this, this paper proposes the following:

Hypothesis 1 (H1).

Analyst logical inconsistency significantly increases stock price crash risk.

2.2. Indirect Impact of Analyst Logical Inconsistency on Stock Price Crash Risk

2.2.1. Financial Risk Channel

Analyst logical inconsistency increases the level of financial risk faced by firms by exacerbating the uncertainty of the external information environment. On the one hand, conflicts among different types of information in analyst reports reduce the market’s degree of consistent interpretation of firm fundamental information, making it difficult for investors to form stable value expectations, thereby weakening the role of analysts as information intermediaries in reducing information asymmetry (M. Chen et al., 2022). When the market’s judgments regarding a firm’s profitability, operational prospects, and risk conditions become differentiated, the uncertainty of the information environment in which the firm operates increases accordingly. In this process, the information conflict characteristics conveyed by logical inconsistency reinforce the market’s uncertain judgment of the firm’s operational stability and future cash flow quality, thereby raising the overall risk exposure of the firm. The Modigliani–Miller theorem states that, under idealized conditions without agency costs, taxes, or market frictions, a firm’s capital structure does not affect its value (Modigliani & Miller, 1958). However, empirical evidence suggests that highly leveraged firms face greater financial risks. Their cost of equity often rises accordingly. This makes the company face greater risks due to leverage, so the risk of sudden decline in stock price becomes greater (Y. Ma et al., 2026). In terms of the stability of a company, when its financial risk is high, its activities are often more uncertain and its profits are easier to change. Due to external shocks or business problems, such firms may experience greater volatility in cash flows, resulting in unexplained fluctuations in their stock prices. Studies have shown that companies with higher financial risks are more likely not to repay their debts, and when external shocks occur, their prices may drop sharply or even collapse (Bae et al., 2021). Based on all this, this paper puts forward the following:

Hypothesis 2 (H2).

Increasing financial risk is an important channel through which analyst logical inconsistency raises stock price crash risk.

2.2.2. Investment Efficiency Channel

By disrupting information dissemination and expectation formation in the capital market, analyst logical inconsistency may reduce corporate investment efficiency. Analysts’ reports not only help investors understand the value of a company, but also change the way companies use their resources through the expected feedback from the market (Healy & Palepu, 2001; Jegadeesh & Kim, 2010). Firms that underinvest fail to fully exploit their available growth opportunities. Therefore, in the long run, they are weaker than other similar companies. However, corporate executives are typically reluctant to acknowledge unfavorable business conditions. Negative information keeps accumulating. When concealed negative news is eventually uncovered by the market, stock prices become vulnerable to sharp adverse corrections and significant adverse movements (Habib & Hasan, 2017). On the contrary, when a company invests too much, it will spend its resources on projects of suboptimal quality or those falling short of the highest standards. Certain executive compensation structures incentivize managers to undertake excessively risky projects, but it may also lead managers to hide their mistakes. They can change or hide information to present the company as remaining on a growth trajectory. This concealment of suboptimal investment behaviors leads to the continuous accumulation over time of bad news and low-quality investment decisions, thereby increasing future stock price crash risk (Benmelech et al., 2010). When a firm experiences underinvestment, its competitiveness compared to other firms in the same industry decreases, and management is also unwilling to disclose such negative information, ultimately leading to stock price crashes (Habib & Hasan, 2017). Although stock-based compensation plans encourage managers to take risks, these plans also prompt them to conceal bad news from suboptimal investment policies to support the illusion of high growth, thus causing bad news and poor investments to accumulate over time, ultimately leading to stock price crashes (Benmelech et al., 2010). Moreover, the correlation between the effectiveness of investments and the likelihood of a stock market collapse can be attributed to the hubris of corporate leaders. Based on this, this paper proposes the following:

Hypothesis 3 (H3).

Reducing investment efficiency is an important channel through which analyst logical inconsistency raises stock price crash risk.

2.2.3. Information Disclosure Channel

Analyst logical inconsistency, by affecting the information interpretation environment and information feedback mechanisms of the capital market, may reduce the overall quality of corporate information disclosure. On the one hand, the contradiction between quantitative predictions and qualitative texts in analyst reports weakens the market’s efficiency in integrating firm-specific information, causing different investors to develop divergent interpretations of the same information signal, thereby reducing the comparability and consistency of information. On the other hand, the existence of logical inconsistency in analyst reports increases the uncertainty of the firm’s information environment, making it difficult for the market to accurately distinguish between changes in firm fundamentals and information expression biases. In corporate management’s information disclosure decisions, they may face a more complex external feedback environment, and their information disclosure behavior is more likely to be influenced by divergent market interpretations, thereby indirectly reducing overall information disclosure quality. Stock price crash risk is closely related to corporate information disclosure quality. Higher information disclosure quality can significantly alleviate information asymmetry between investors and firms, thereby reducing stock price crash risk at multiple levels. Through the lens of information asymmetry framework, when corporate governance is weak or the external environment is imperfect, management can hide negative information by choosing what to say or delaying news. In this way, they can better control the company information and transfer the risk to the management and external market to some extent (Jin & Myers, 2006). As this state of information opacity persists and accumulates, when negative information reaches a certain level, the market often finds it difficult to digest the related risks in a timely manner, leading to the concentrated release of information over a short period, triggering sharp stock price declines or even crashes (Hutton et al., 2009). Therefore, when information disclosure quality is higher, the market can obtain the firm’s true information more timely and adequately, thereby reducing valuation biases caused by information asymmetry and effectively suppressing the hidden accumulation of negative information. Based on this, this paper proposes the following:

Hypothesis 4 (H4).

Reducing information disclosure quality is an important channel through which analyst logical inconsistency raises stock price crash risk.

3. Methodology and Data

3.1. Econometric Models

Drawing from the theoretical analysis outlined above, this chapter develops the following econometric model to investigate how analyst logical inconsistency influences the stock price crash risk:

$$NCSKE{W}_{i,t+1}={\alpha }_0+{\alpha }_{1}INCO{N}_{it}+{\displaystyle {\sum }_{j=1}^k{\beta }_j{X}_{ijt}}+{\gamma }_i+{\eta }_t+{\epsilon }_{it}$$

(1)

$$DUVO{L}_{i,t+1}={\alpha }_0+{\alpha }_{1}INCO{N}_{it}+{\displaystyle {\sum }_{j=1}^k{\beta }_j{X}_{ijt}}+{\gamma }_i+{\eta }_t+{\epsilon }_{it}$$

(2)

where i and t denote firm and year, $NCSKE{V}_{i,t+1}$ and $DUVO{L}_{i,t+1}$ denote the stock price crash risk, $INCO{N}_{it}$ denotes analyst logical inconsistency, and X denotes the vector of control variables, $X_{ijt}$, including firm age, firm size, liquidity ratio, market value, leverage ratio, return on assets, cash flow intensity, operating margin, ownership concentration, CEO duality, separation of ownership and control, proportion of independent directors, and audit quality. ${\gamma }_i$ and ${\eta }_t$ represent firm fixed effects and year fixed effects, and ${\epsilon }_{it}$ is the random error term.

3.2. Variables

3.2.1. Dependent Variable: Stock Price Crash Risk (NCSKEW and DUVOL)

The dependent variable in this paper is stock price crash risk. Following Hutton et al. (2009), we calculate the negative conditional skewness (NCSKEW) and the down-to-up volatility ratio (DUVOL) based on Equation (3) to measure stock price crash risk.

$$r_{it}={\beta }_0+{\beta }_{1i}{r}_{m,t-2}+{\beta }_{2i}{r}_{m,t-1}+{\beta }_{3i}{r}_{mt}+{\beta }_{4i}{r}_{m,t+1}+{\beta }_{5i}{r}_{m,t+2}+{\epsilon }_{it}$$

(3)

where $r_{it}$ is the weekly return of stock i in week t; $r_{mt}$ is the value-weighted average return of all A-shares in week t; and the remaining terms are lagged and lead terms. The firm-specific weekly return adjusted for market returns is used to calculate NCSKEW and DUVOL as follows:

$$NCSKE{W}_{it}=-\left[n{\left(n-1\right)}^{3/2}{\displaystyle \sum W_{it}^3}\right]{\displaystyle /}\left[\left(n-1\right)\left(n-2\right){\left({\displaystyle \sum W_{it}^2}\right)}^{3/2}\right]$$

(4)

$$DUVO{L}_{it}=ln\left\{\left[\left(n_u-1\right){\displaystyle {\sum }_{down}{W}_{it}^2}\right]{\displaystyle /}\left[\left(n_d-1\right){\displaystyle {\sum }_{up}{W}_{it}^2}\right]\right\}$$

(5)

where $W_{it}$ is the market-adjusted weekly return of stock i in week t; $n$ is the number of weeks, and $n_u$ and $n_d$ are the number of weeks in which the firm-specific weekly return $W_{it}$ is greater than and less than the annual mean, respectively. Larger values of NCSKEW and DUVOL indicate greater stock price crash risk (Hutton et al., 2009). The method employed here is the comprehensive market equal-weighted method.

3.2.2. Independent Variable: Analyst Logical Inconsistency (INCON)

This chapter follows the research of L. J. Ma et al. (2022) to construct a deviation index system from three aspects—valuation conclusions, textual features, and logical consistency—using multi-level quantitative indicators to deeply examine analyst logical inconsistency.

Regarding valuation conclusions, we focus on the degree of deviation between analysts’ earnings forecasts and the market consensus expectation, as shown in Equation (6):

$$BiasE{F}_{ikt}={\displaystyle \frac{E{F}_{ikt}-ConsensusE{F}_{it}}{ConsensusE{F}_{it}}}$$

(6)

where $k$ denotes an analyst report, $E{F}_{ikt}$ is the analyst’s earnings forecast, and $ConsensusE{F}_{it}$ is the mean of all effective analyst forecasts prior to the report’s issuance. When $BiasE{F}_{ikt}>0$, it indicates that the analyst’s earnings forecast is relatively optimistic compared to the market consensus, conveying new positive information; when $BiasE{F}_{ikt}<0$, it indicates that the analyst holds a pessimistic attitude towards the firm’s future profitability, conveying new negative information. The greater the magnitude of $BiasE{F}_{ikt}$. A larger absolute value indicates a greater deviation from the market consensus, although such a deviation may reflect either incremental information or forecast noise.

Regarding textual features, based on the large language model, we interpret the degree of tone deviation in the text after measuring analyst sentiment (Xing, 2025).

Step 1: Using Python3.11 web crawler technology, we collected over 300,000 analyst reports related to Chinese A-share listed companies from Eastmoney1 for the period 2016 to 2023. From these, we extracted the report publication date, title, analyst name and affiliated institution, quantitative forecast data, and the summary text content. To ensure data quality and relevance, the collected reports were processed as follows: (1) we excluded macroeconomic and industry research reports, thereby retaining only sell-side analyst research; (2) we excluded reports that could not be clearly matched to a specific listed company; (3) we excluded reports that were associated with multiple listed companies simultaneously; (4) to ensure that the impact of analyst reports on stock prices was not disturbed by other performance information, we further excluded non-earnings reports such as in-depth reports and corporate research reports, ultimately obtaining 40,135 reports. All reports were segmented by periods and semicolons to create the sentence database to be predicted.

Step 2: All reports were segmented at the sentence level to construct a text column L for each report. In the tokenization process, Jieba2 was used to filter out mood words, stop words, and other meaningless content. Combined with tokenization methods, sentence-level text vectors were formed. According to different subtasks, the frequency of each word in the vector was used to convert the text vector into the corresponding frequency vector Norm. We employed a Chinese financial text sentiment dictionary. Using this dictionary, textual sentiment indicators for the Chinese stock market were constructed to measure changes in investor and market sentiment. Textual sentiment values range between −1 and 1. When the positive sentiment score (Pos) surpasses the negative sentiment score (Neg), the sentiment is classified as positive, with values nearing 1 reflecting a fairly favorable outlook. The study considered the impact of negative words (e.g., “not”, “no”, “without”, “not at all”, “impossible”, etc., totaling 26 words) on sentiment polarity.

Step 3: For previous deep learning models, we employed the Chinese BERT sentiment classification model Cemotion2.03 to obtain the sentiment score for each sentence (Gardazi et al., 2025), setting thresholds to classify sentences with scores > 0.9 as positive sentiment sentences, and those with scores < 0.01 as negative sentiment sentences.

$$\begin{array}{l}Po{s}_{bert}=Bert\_Score>0.9\\ Ne{g}_{bert}=Bert\_Score<0.01\end{array}$$

(7)

For large model sentiment classification, we employed the open source Qwen1.5-14B-Chat model, using the following prompt to enable the model to output sentiment classification scores for each sentence:

“You are a financial analyst. You are given the following text: {context}. Please classify the sentiment contained in the above text. Positive sentiment is 1, neutral sentiment is 0, and negative sentiment is −1. Output in the following format: {context: score}.”

To construct the analyst sentiment dataset, we randomly selected 5300 sentences from the full dataset for annotation, providing materials for subsequent large model fine-tuning. Fine-tuning large models is essentially a machine learning process. After the general model learns basic patterns during pre-training, fine-tuning can be employed, using specific smaller datasets for additional training, which are usually associated with specific tasks. The purpose of fine-tuning is to adjust the parameters of the pre-trained model to better adapt to the requirements of specific tasks. However, due to the inherent characteristics of large models, the construction method differs from previous deep learning training datasets. The primary difference is that the fine-tuning dataset for large models can train multiple subtasks simultaneously in stages.

Following the sentiment classification methods mentioned above, we annotated the analyst texts, classifying sentences into three categories: positive, neutral, and negative. When we annotate emotions in the text of the analysis report, we follow the following rules: Sentences describing successful operations, managerial efforts and achievements, or optimistic expectations were labeled as positive. Text that only explains operational and financial facts or makes an objective evaluation of the company is marked as neutral. Texts that mainly describe a very difficult operating environment, worrying development prospects, or express negative evaluation of management’s operational performance are marked as negative emotions. This kind of classified annotation is helpful to understand and analyze the emotional trend in the report more accurately, thus providing investors with more accurate baseline information. The format of training data is as follows:

{Input: As a financial analyst, please tell us which of the following options belongs to the emotion in the text. With the recovery of consumer electronics power industry and the increase of new energy power industry capacity, we predict that the revenue growth rate of the company’s power industry will be 301%, 38% and 19% respectively from 2023 to 2025, and the gross profit margin will be 10.0%, 16.0% and 16.7% respectively. Options: A. Positive B. Neutral C. Negative}

The experiment was carried out on the Xeon-6330 X86 server based on Linux, with NVIDIA A40-48G GPU and 120 GB of RAM. The whole training lasted about two hours. In order to improve the results of the model, we used a set of adjusted parameters and driving strategies. First of all, we selected the open source LLaMA-Factory training framework to experiment. This framework not only provides a complete training process management, but also has many tools to optimize the model. For the optimizer, we choose Adam optimizer because its adaptive learning rate adjustment mechanism is particularly suitable for solving optimization problems with smooth gradient and non-stationary objectives. The learning rate was set to 1 × 10⁻⁵, which is a very cautious setting and helps to ensure stable convergence without excessively adjusting parameters.

To improve training efficiency under limited computational resources, we used the gradient accumulation strategy and set the gradient accumulation step to 2. This strategy enables us to obtain the benefits of a larger effective batch size during training, while greatly reducing memory usage. In order to adjust the stability of training, we introduced NEFTune technology and set the NEFTune noise alpha parameter to 5. This parameter is selected through many experiments, and we find it is the best: it improves the robustness of the model and helps to avoid overfitting.

The model was trained for three epochs, with each epoch consisting of one complete pass through the training set. This setting allows the model to learn the features in the data set well, and avoids performance degradation due to too long training time. As for the calculation accuracy, we choose the calculation type of bfloat16(bf16), which is a calculation scheme with mixed accuracy and achieves a good balance between numerical stability and calculation efficiency. Compared with the traditional float32, bf16 greatly reduces the memory usage (about 50%) and has a dynamic range close to float32. Therefore, it is suitable for training large language models. The combination of these parameters and strategies helped us achieve an ideal balance among model performance, training efficiency, and computational resource utilization. Further, based on analyst sentiment sentences, we constructed the degree of tone deviation in analyst texts as follows:

$$BiasTon{e}_{ikt}={\displaystyle \frac{Po{s}_{ikt}-Ne{g}_{ikt}}{Sentence{s}_{ikt}}}$$

(8)

where $Po{s}_{ikt}$ and $Ne{g}_{ikt}$ represent the number of positive and negative sentiment sentences in analyst reports, and $Sentenc{e}_{ikt}$ is the total number of sentences in the report. $BiasTon{e}_{ikt}$ effectively captures the analyst’s perspective tendency when describing company development, industry trends, and market direction. When $BiasTon{e}_{ikt}>0$, it indicates that the textual expression is relatively optimistic compared to the capital market, conveying new positive information; when $BiasTon{e}_{ikt}<0$, it indicates that the report tone is relatively pessimistic, conveying new negative information.

This paper considers that only when there is a contradiction between the analyst’s quantitative forecast and qualitative analysis—that is, when the degree of deviation between earnings forecast and tone interpretation exceeds a certain threshold $\theta$—does the report contain logically inconsistent information. Here we set $\theta$ to 0. If $BiasE{F}_{ikt}>\theta ,BiasTon{e}_{ikt}<-\theta$ or $BiasE{F}_{ikt}<-\theta ,BiasTon{e}_{ikt}>\theta$, then the analyst logical inconsistency of that report $INCO{N}_{ikt}=1$, reflecting report quality issues and potential conflicts of interest. Finally, we sum the $INCO{N}_{ikt}$ for all analyst reports of a firm to calculate the firm’s analyst logical inconsistency $INCO{N}_{it}$.

3.2.3. The Control Variables

We mainly select the following 13 control variables, with specific definitions as shown in

Table 1. Definitions of control variables.

Variables	Measure
lnAge	Natural logarithm of (current year − listing year + 1)
lnSize	Natural logarithm of total assets
Liquid	Current assets/total assets
TobinQ	Tobin’s Q value
Leverage	Total liabilities/total assets
ROA	Net profit/total assets
Cash	Cash and cash equivalents at year-end/total assets
Profit	Operating profit/operating revenue
Share	Aggregate stakeholder ownership proportions among the top five investors
Duality	Takes the value of 1 when the chairperson and CEO are identical, and 0 otherwise.
Separate	Difference between ultimate controller’s control rights and ownership rights
Director	Number of independent directors at year-end/total board size
Audit	Equals 1 if the annual report is audited by one of the top 10 domestic accounting firms, otherwise 0

.

3.3. Data

The study employs A-share listed firms spanning 2016–2023 as the study population, employing these selection standards: (1) exclude ST and PT firms4; (2) exclude firms in the financial sector; and (3) winsorize all continuous variables at the 1st and 99th percentiles. After the above processing, the final sample includes 3784 listed companies with 12,460 firm-year observations. The data in this paper are sourced from the iFind database, Wind database, CNRDS database, and CSMAR database. The data and code in the empirical analysis can be found in the Supplementary Materials.

The calculation results of the analyst logical inconsistency constructed in this paper are consistent with those of existing studies (L. J. Ma et al., 2022), and the text sentiment indicators constructed using large language models can better capture the deep background and complexity of human language and can better understand the context and semantic relationships through datasets (Albladi et al., 2025). Furthermore, to ensure the validity of the regression results, multicollinearity tests were conducted for each explanatory variable. The findings displayed in

Table 2. Descriptive statistics and multicollinearity diagnostics.

Variables	N	Mean	SD	Min	Median	Max	VIF	1/VIF
NCSKEW	12,460	−0.2726	0.6344	−2.1996	−0.2205	1.4931	-	-
DUVOL	12,460	−0.1928	0.4289	−1.2392	−0.1717	0.8936	-	-
INCON	12,460	1.4327	1.8633	0	1	16	1.13	0.8874
lnAge	12,460	2.9812	0.3078	2.0794	2.9957	3.6109	1.13	0.8864
lnSize	12,460	22.6958	1.3506	20.4198	22.4823	26.8499	1.87	0.5345
Liquid	12,460	0.5851	0.1983	0.1075	0.5997	0.9517	1.40	0.7150
TobinQ	12,460	2.1311	1.3438	0.8411	1.7099	8.5446	1.29	0.7760
Leverage	12,460	0.4012	0.1908	0.0538	0.3980	0.8328	1.95	0.5133
ROA	12,460	0.0568	0.0559	−0.1556	0.0531	0.2280	2.93	0.3411
Cash	12,460	0.1740	0.1283	0.0158	0.1368	0.6326	1.50	0.6675
Profit	12,460	0.1266	0.1495	−0.5090	0.1124	0.5753	2.75	0.3636
Share	12,460	0.3356	0.1493	0.0843	0.3125	0.7498	1.12	0.8943
Duality	12,460	0.3506	0.4772	0	0	1	1.09	0.9203
Separate	12,460	0.0450	0.0704	0.0000	0.0007	0.2815	1.05	0.9547
Director	12,460	0.3791	0.0541	0.3333	0.3636	0.5714	1.02	0.9758
Audit	12,460	0.6612	0.4733	0	1	1	1.03	0.9740

demonstrate that the variance inflation factors (VIFs) for every explanatory variable sit comfortably below the 3-point mark, well within the desirable limit. Additionally, the inverse VIF for each variable registers over the 0.3 benchmark, pointing to a modest degree of multicollinearity among the explanatory factors. There is no severe multicollinearity among the explanatory variables selected in this paper, and the unique information of each variable is effectively retained. The model specification is reasonable, allowing for further baseline regression analysis.

4. Empirical Results

4.1. Regression Results

Table 3. Baseline regression results.

Variables	NCSKEW	DUVOL	NCSKEW	DUVOL
INCON	0.0172 *** (0.0036)	0.0128 *** (0.0026)	0.0123 *** (0.0037)	0.0095 *** (0.0026)
Constant	−0.2102 *** (0.0197)	−0.1525 *** (0.0136)	−0.6453 (0.8081)	−0.0682 (0.5522)
Controls	No	No	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.0338	0.0311	0.0403	0.0372

Sig. *** p < 0.01. Parentheses contain firm-level clustered standard errors.

presents the baseline regression results obtained from the two-way fixed-effects models. Constant represents the constant term; Controls indicates whether control variables are included in the model; Firm FE indicates whether to include firm fixed effects in the model; and Year FE indicates whether to include year fixed effects in the model. The results in the following table are consistent with this statement. Therefore, no further elaboration is necessary. Columns (1) and (2) use NCSKEW and DUVOL as the dependent variables, respectively, without including the control variables. The coefficients on INCON are positive and statistically significant at the 1% level in both columns. After the control variables are included, the estimated coefficients on INCON remain positive and statistically significant at the 1% level, with coefficients of 0.0123 and 0.0095 in Columns (3) and (4), respectively. It is worth noting that the R² value in this article is relatively low. This is because the factors influencing the stock market crash are extremely complex and include a large amount of market sentiment, macroeconomic fluctuations, and company-specific information that cannot be fully quantified. The two-way fixed effects model used in this paper has eliminated a large amount of systematic variation caused by the inherent characteristics of the companies and the macro time trends by including company and year dummy variables. The explanatory power of the core independent variable for the within-group variation in the dependent variable is much lower than the total variation. Therefore, a low R² is a common result of the fixed effects model and does not indicate that the model is invalid. These results are consistent with Hypothesis 1.

4.2. Endogeneity Issues

4.2.1. Two-Stage Least Squares

To mitigate potential endogeneity problems caused by reverse causality and omitted variables, this chapter employs a “dummy variable indicating whether the value is above the industry mean” as an instrumental variable for analyst logical inconsistency. Specifically, if a firm’s analyst logical inconsistency indicator is higher than the industry-year mean, it takes the value of 1; otherwise, it is 0. Regarding relevance, analyst behavior within an industry exhibits significant peer group characteristics and information spillover effects. The report styles, information expression methods, and logical consistency levels of different firms are jointly influenced by the industry information environment and analyst coverage structure. Whether a firm is above the industry average effectively reflects the relative strength of its analyst logical inconsistency, thus meeting the exogeneity condition necessary for an instrumental variable. In terms of exogeneity, this variable is primarily constructed based on relative classification at the industry level, with changes mainly originating from the overall industry information environment and analyst group behavioral characteristics, rather than shocks to the firm’s individual stock price crash risk. The industry mean grouping variable itself does not directly affect the process of accumulation of future negative information and price crash mechanisms; therefore, this instrumental variable can satisfy the exogeneity assumption to a certain extent. The first-stage results in

Table 4. Regression results of two-stage least squares.

Variables	INCON	NCSKEW	INCON	DUVOL
INCON		0.0116 ** (0.0054)		0.0089 ** (0.0037)
IV	2.6135 *** (0.0255)		2.6135 *** (0.0255)
Kleibergen–Paap rk LM	1418.96		1418.96
Kleibergen–Paap rk Wald F	10,057.72 [16.38]		10,057.72 [16.38]
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
N	11,526	11,526	11,526	11,526

Sig. *** p < 0.01, and ** p < 0.05. Parentheses contain firm-level clustered standard errors. Values in square brackets are the critical values at the 10% level of the Stock–Yogo weak instrument identification F-test.

show that the instrumental variable (above-industry-mean dummy) is highly positively correlated with INCON, satisfying the relevance requirement of the instrumental variable. At the same time, the LM value and Wald F value for weak instrument testing are 1418.96 and 10,057.72, respectively, both exceeding the critical values, rejecting the weak-instrument null hypothesis and further confirming the rationality of the chosen instrumental variable. The estimated coefficients on INCON are positive and statistically significant at the 5% levels, respectively, indicating that after accounting for potential endogeneity arising from reverse causality and omitted variables, the conclusion that analyst logical inconsistency increases stock price crash risk remains valid.

4.2.2. Propensity Score Matching

To mitigate observable selection differences between firms with and without analyst logical inconsistency, we employ propensity score matching. First, firms with analyst logical inconsistency levels greater than 0 are designated as the treatment group, and firms with analyst logical inconsistency levels equal to 0 are designated as the control group. Matching is performed using a combination of caliper matching and nearest-neighbor matching. The caliper setting is dialed in at 0.01, and the closest match ratios are locked at 1:1 and 1:2. When it comes to the 1:1 nearest-neighbor matches with the 0.01 caliper, the average treatment effects on the treated (ATT) register at 0.0681 and 0.0437, respectively, and these figures are statistically significant at the 1% level. This indicates that the NCSKEW and DUVOL values of listed firms with analyst logical inconsistency are, on average, about 0.0681 and 0.0437 higher, respectively. The standardized biases of all covariates are less than 10%, indicating good sample balance. In the findings from a 1:2 nearest-neighbor matching procedure using a caliper of 0.01, the average treatment effect on the treated (ATT) stands at 0.0782 and 0.0529, respectively, both of which are statistically significant at the 1% level. This indicates that the NCSKEW and DUVOL values of listed firms with analyst logical inconsistency are, on average, about 0.0782 and 0.0529 higher, respectively, and the standardized biases of all covariates remain less than 10%, indicating good sample balance. Based on the two matching methods, we further re-run regressions on the matched samples.

Table 5. Regression results of propensity score matching.

Variables	NCSKEW	DUVOL	NCSKEW	DUVOL
INCON	0.0153 *** (0.0057)	0.0111 *** (0.0043)	0.0170 *** (0.0046)	0.0121 *** (0.0033)
Constant	−0.1387 (1.2679)	−0.0920 (0.8988)	0.0325 (0.9957)	0.2907 (0.6846)
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.0454	0.0405	0.0412	0.0372
N	6247	6247	8943	8943

Sig. *** p < 0.01. Parentheses contain firm-level clustered standard errors.

presents the outcomes, with INCON’s coefficients registering a statistically significant positivity at the 5% and 1% confidence intervals. After mitigating selection bias through propensity score matching, the conclusion that analyst logical inconsistency increases stock price crash risk remains unchanged.

4.3. Robustness Checks

4.3.1. Replacing the Core Independent Variable

To rule out the influence of alternative measurement methods, this paper further adjusts the identification threshold for analyst logical inconsistency to 5%, 10%, 15%, and 20%, respectively, re-constructs the analyst logical inconsistency indicators, and re-runs the regression analysis. When the consistency of the direction and magnitude of the regression coefficients for the analyst logical inconsistency persists across various threshold levels, it suggests that the findings of the study are not contingent on the particular choice of threshold and are thus robust. The robustness of the results, with the core independent variable being substituted, is illustrated in

Table 6. Robustness test results of replacing the core independent variable.

Variables	NCSKEW	NCSKEW	NCSKEW	NCSKEW	DUVOL	DUVOL	DUVOL	DUVOL
INCON05	0.0154 *** (0.0046)				0.0108 *** (0.0033)
INCON10		0.0126 ** (0.0059)				0.0112 *** (0.0042)
INCON15			0.0164 ** (0.0077)				0.0158 *** (0.0055)
INCON20				0.0274 *** (0.0098)				0.0221 *** (0.0072)
Constant	−0.6416 (0.8081)	−0.6760 (0.8095)	−0.6847 (0.8095)	−0.6755 (0.8097)	−0.0717 (0.5525)	−0.0861 (0.5532)	−0.0903 (0.5531)	−0.0901 (0.5529)
Controls	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
R2	0.0403	0.0396	0.0396	0.0399	0.0369	0.0365	0.0366	0.0367
N	12,460	12,460	12,460	12,460	12,460	12,460	12,460	12,460

Sig. *** p < 0.01, and ** p < 0.05 Parentheses contain firm-level clustered standard errors.

. Regardless of the chosen threshold, the coefficient for the core independent variable remains notably positive, signifying that the impact of analyst logical inconsistency on the risk of stock price crashes is unaffected by variations in the initial threshold.

4.3.2. Replacing the Dependent Variable

Market return calculations differ mainly in portfolio construction and weighting methods. Different methods have different emphases on characterizing overall market returns, which may affect the calculation of individual stock weekly specific returns and stock price crash risk indicators. In the robustness test, we further re-calculate individual stock weekly specific returns using the sub-market value-weighted average method, sub-market total market value-weighted average method, sub-market equal-weighted average method, comprehensive market value-weighted average method, and comprehensive market total market value-weighted average method, and re-construct the NCSKEW and DUVOL indicators accordingly.

Table 7. Robustness test results of replacing the dependent variable.

Variables	NCSKEW	DUVOL	NCSKEW	DUVOL	NCSKEW	DUVOL	NCSKEW	DUVOL	NCSKEW	DUVOL
	Sub-Market Value-Weighted		Sub-Market Total Value-Weighted		Sub-Market Equal-Weighted		Comprehensive Value-Weighted		Comprehensive Total Value-Weighted
INCON	0.0113 *** (0.0040)	0.0091 *** (0.0027)	0.0111 *** (0.0040)	0.0075 *** (0.0027)	0.0124 *** (0.0039)	0.0091 *** (0.0027)	0.0100 ** (0.0039)	0.0081 *** (0.0027)	0.0099 ** (0.0039)	0.0072 *** (0.0027)
Constant	−1.2129 (0.8578)	−0.3722 (0.5727)	−1.0381 (0.8642)	−0.7424 (0.5766)	−1.0564 (0.8435)	−0.3722 (0.5727)	−1.0147 (0.8414)	−0.5286 (0.5632)	−0.8497 (0.8554)	−0.3397 (0.5722)
Controls	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
R2	0.0350	0.0405	0.0353	0.0354	0.0407	0.0405	0.0348	0.0356	0.0342	0.0338
N	12,460	12,460	12,460	12,460	12,460	12,460	12,460	12,460	12,460	12,460

Sig. *** p < 0.01, and ** p < 0.05 Parentheses contain firm-level clustered standard errors.

delineates the regression outcomes linking analyst logical inconsistency to NCSKEW and DUVOL. The coefficients on INCON remain positive and statistically significant at the 1% or 5% level across all alternative measures of stock price crash risk. When we alter the computation technique for the dependent variable, our theory about the increase in stock price crash risk due to analyst logical inconsistency stands firm. This association remains unaffected by either the composition of the market portfolio or the weighting schemes applied.

4.3.3. Considering Potential Influencing Factors

In order to definitively dismiss the impact of any potential omitted variables on the study’s findings, the paper undertakes robustness analyses by introducing additional control variables that could potentially sway the risk of stock market crashes. Referring to existing research, on the basis of the baseline regression model, we sequentially include firm-specific characteristic variables, including stock average return (Ret), stock return volatility (Sigma), book-to-market ratio (Bm), and additionally include analyst coverage (Analyst) as a control variable for re-regression. Specifically, stock average return reflects the firm’s stock return level and affects stock price crash risk; stock return standard deviation measures stock price volatility, and greater volatility increases the likelihood of future extreme negative return events; and the book-to-market ratio reflects the firm’s growth characteristics and value attributes, and firms at different growth stages have significantly different information environments and risk levels, which may affect stock price crash risk. At the same time, analyst coverage reflects the degree of information attention and external supervision intensity in the capital market. Higher analyst coverage typically implies higher corporate information transparency and lower information asymmetry, which may affect stock price crash risk. The regression results after separately controlling for these potential influencing factors are shown in

Table 8. Robustness test results considering potential influencing factors.

Variables	NCSKEW	DUVOL	NCSKEW	DUVOL	NCSKEW	DUVOL	NCSKEW	DUVOL
INCON	0.0109 *** (0.0037)	0.0086 *** (0.0026)	0.0109 *** (0.0037)	0.0087 *** (0.0026)	0.0083 ** (0.0037)	0.0070 ** (0.0027)	0.0110 *** (0.0039)	0.0089 *** (0.0029)
Ret	2.3556 *** (0.5581)	1.5789 *** (0.3437)
Sigma			1.2900 *** (0.2452)	0.7805 *** (0.1663)
Bm					−0.5610 *** (0.0739)	−0.3535 *** (0.0513)
Analyst							0.0009 (0.0010)	0.0005 (0.0007)
Constant	−1.2428 (0.8182)	−0.4686 (0.5567)	−1.3338 (0.8159)	−0.4847 (0.5570)	−0.3682 (0.8077)	0.1065 (0.5529)	−0.4106 (0.8360)	0.0916 (0.5742)
Controls	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
R2	0.0438	0.0405	0.0430	0.0392	0.0463	0.0421	0.0405	0.0373
N	12,460	12,460	12,460	12,460	12,460	12,460	12,290	12,290

Sig. *** p < 0.01, and ** p < 0.05 Parentheses contain firm-level clustered standard errors.

. The coefficients for INCON still show a significant positive effect, further affirming the stability of the findings.

4.3.4. Changing the Sample Period

To further rule out the impact of major event shocks in special years on the empirical results, this paper conducts robustness tests by changing the sample period. Considering that the stock price crash risk indicator uses a one-year lag, the 2016 sample corresponds to the post-2015 A-share market turbulence adjustment period. During this phase, the capital market was continuously affected by factors such as severe stock market fluctuations, changes in market sentiment, and regulatory policy adjustments. Corporate stock price volatility may have been more driven by macro-market factors than by the firm’s own information characteristics. Therefore, this chapter excludes the 2016 sample and re-runs the regressions to reduce the interference of abnormal market fluctuations on the research conclusions. In

Table 9. Robustness test results of changing the sample period.

Variables	NCSKEW	DUVOL	NCSKEW	DUVOL
INCON	0.0103 ** (0.0041)	0.0086 *** (0.0029)	0.0128 *** (0.0038)	0.0098 *** (0.0028)
Constant	−0.3282 (0.8486)	−0.0247 (0.5811)	−0.1443 (0.9398)	0.1629 (0.6498)
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.0439	0.0407	0.0425	0.0393
N	10,925	10,925	11,186	11,186

Sig. *** p < 0.01, and ** p < 0.05 Parentheses contain firm-level clustered standard errors.

, Columns (1) and (2) present the test findings. In addition, in 2020, due to the impact of sudden public events such as the COVID-19 pandemic, corporate operations, analyst information acquisition, and capital market functioning were all significantly affected. Stock price crash risk may have been significantly influenced by exogenous systemic risks, thereby weakening the true economic effects of analyst logical inconsistency. Therefore, this paper further excludes the 2020 sample for re-regression to test whether the research conclusions are affected by special events such as the pandemic shock. In Table 9, Columns (3) and (4) display the regression outcomes. The regression results indicate that, after excluding the 2016 and 2020 samples separately, the direction and significance of the regression coefficients of analyst logical inconsistency do not change substantially, indicating that the research conclusions are not driven by systemic shocks in special years and exhibit good robustness.

4.3.5. Considering Other Fixed Effects

To bolster the credibility of the research findings and mitigate the impact of unseen industry and local variables, the paper incorporates industry-level and provincial-level fixed effects in the robustness analysis. On the one hand, different industries differ significantly in terms of business models, market competition environments, regulatory systems, and risk characteristics. These industry-level heterogeneities may simultaneously affect analyst logical inconsistency and corporate stock price crash risk. By including industry fixed effects, we can effectively control for unobservable time-invariant industry-level factors, reducing the interference of industry characteristics on the estimation results. On the other hand, different regions differ significantly in terms of economic development levels, financial market maturity, legal environments, information disclosure systems, and investor protection levels. These regional characteristics may affect analysts’ information acquisition and transmission efficiency, while also influencing the corporate information environment and the formation of stock price crash risk. Therefore, further controlling for province fixed effects helps eliminate the influence of unobservable regional factors on the research conclusions, improving the reliability of the estimation results. The test results are shown in

Table 10. Robustness test results of considering other fixed effects.

Variables	NCSKEW	DUVOL	NCSKEW	DUVOL
INCON	0.0122 *** (0.0037)	0.0094 *** (0.0026)	0.0122 *** (0.0037)	0.0095 *** (0.0026)
Constant	−0.0395 (0.8654)	0.1082 (0.5866)	−0.3550 (0.8344)	0.0401 (0.5785)
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
Industry FE	Yes	Yes	-	-
Province FE	-	-	Yes	Yes
R2	0.0417	0.0389	0.0439	0.0404
N	12,460	12,460	12,460	12,460

Sig. *** p < 0.01. Parentheses contain firm-level clustered standard errors.

. Regardless of whether industry fixed effects or province fixed effects are added, the consistency in the coefficients, which measure the logic flaws among analysts, remains strongly favorable even when examined at the 1% significance level, conclusively affirming the dependability of the original model.

4.4. Heterogeneity Analysis

4.4.1. News Media Attention

As an important information intermediary in the capital market, news media not only undertakes the functions of corporate information dissemination and supervisory governance but also affects the market’s interpretation and pricing efficiency of analyst reports through channels such as information diffusion, public opinion guidance, and investor attention. This chapter uses the median of news media coverage as the classification criterion, dividing the sample into high news media attention and low news media attention groups for heterogeneity analysis.

Table 11. Heterogeneity test results based on news media attention.

Variables	NCSKEW	DUVOL	NCSKEW	DUVOL
	High Media Attention		Low Media Attention
INCON	0.0183 *** (0.0050)	0.0144 *** (0.0037)	0.0089 (0.0067)	0.0064 (0.0047)
Constant	−1.7563 (1.1386)	−0.8010 (0.7932)	0.6596 (1.5152)	1.0448 (1.0099)
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.0504	0.0466	0.0423	0.0400
N	6205	6205	6255	6255

Sig. *** p < 0.01. Parentheses contain firm-level clustered standard errors.

displays the findings. Regression outcomes for the group with substantial news media coverage appear in Columns (1) and (2). The coefficients of analyst logical inconsistency are 0.0183 and 0.0144. Columns (3) and (4) present the regression outcomes for the group with minimal news media coverage. The coefficients indicating analyst logical inconsistency failed to reach statistical significance at the 1% level. It shows that the impact of analyst logical inconsistency on stock price crash risk is primarily concentrated in firms with high news media attention. When news media attention is high, the media can accelerate the dissemination of analyst reports, increasing investor attention to analyst views, making logically inconsistent analyst reports more likely to influence market expectations and exacerbate investor misjudgments of the firm’s true operating conditions. At the same time, in the process of reposting and interpreting analyst views, news media may further amplify information noise and market sentiment, causing the market to react more strongly to logically inconsistent information, thereby increasing the risk of stock price crashes when negative information is centrally released.

4.4.2. Corporate Reputation

Corporate reputation, as an important intangible asset accumulated over long-term business operations, not only enhances investor attention to firm information but also affects the market’s interpretation and pricing of external information. Firms with higher reputations typically have higher market attention and a broader investor base, making all types of information they release more likely to attract capital market attention. The information conveyed by analyst reports is also more likely to influence investor expectations. Conversely, firms with lower reputations have limited market attention, and the informational influence of analyst reports is relatively smaller. Therefore, this chapter uses the median of corporate reputation as the classification criterion, dividing the sample into high reputation and low reputation groups for heterogeneity analysis. The data are presented in

Table 12. Heterogeneity test results based on corporate reputation.

Variables	NCSKEW	DUVOL	NCSKEW	DUVOL
	High Reputation		Low Reputation
INCON	0.0119 ** (0.0047)	0.0089 *** (0.0034)	0.0012 (0.0073)	0.0031 (0.0052)
Constant	−2.5120 ** (1.1754)	−1.5559 * (0.8445)	0.6989 (1.4056)	0.7122 (0.9577)
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.0591	0.0526	0.0344	0.0306
N	6230	6230	6230	6230

Sig. *** p < 0.01, ** p < 0.05, and * p < 0.10. Parentheses contain firm-level clustered standard errors.

. This study follows the approach of Ran and Ji (2024) and selects 12 indicators for evaluating corporate reputation, such as asset, revenue, net profit, and industry-specific rankings, as well as leverage, current, and long-term debt ratios, earnings and dividend per share, whether the company has been audited by a Big Four firm, the sustainable growth rate, and the composition of independent directors. The enterprise reputation scores are calculated using the factor analysis method for these 12 indicators. Among them, the principal component method is used to calculate the factor loadings, and orthogonal rotation is performed on the factors. Columns (1) and (2) display the regression outcomes for the high-reputation cohort, with coefficients for analyst logical inconsistency at 0.0119 and 0.0089, respectively, and statistical significance at the 5% and 1% levels. Columns (3) and (4) present the corresponding results for the low-reputation group. The regression coefficients of analyst logical inconsistency are not significant. This indicates that the promoting effect of analyst logical inconsistency on stock price crash risk is mainly concentrated in firms with higher reputations. The reason may be that firms with better reputations are subject to continuous attention from investors, media, and analysts, and analyst reports have stronger information dissemination effects and market influence. Based on the research presented in this article, when analyst reports are logically contradictory, investors are more likely to make wrong expectations for these reports (Kothari et al., 2016), thus increasing the noise of market information and exacerbating stock mispricing (Silva & Thesmar, 2024). When all the hidden negative information appears together, the stock price reaction is very severe, and the risk of stock price collapse is higher.

4.4.3. ESG Level

Companies with good ESG performance usually have higher information transparency, stronger internal governance structure and more effective stakeholder control. All these are helpful to reduce information asymmetry and opportunism of managers. In contrast, companies with low ESG level usually have poor information disclosure quality and weak governance mechanism. Therefore, they are more likely to accumulate negative information, so the impact of analyst logical inconsistency on the market becomes greater. Therefore, this chapter divides the samples into two groups—high ESG level and low ESG level—for heterogeneity analysis. Companies with an ESG rating of A, AA or AAA are rated as high ESG, and all other companies are considered as low ESG. The results of regression are shown in

Table 13. Heterogeneity test results based on ESG level.

Variables	NCSKEW	DUVOL	NCSKEW	DUVOL
	High ESG Level		Low ESG Level
INCON	0.0127 (0.0116)	0.0060 (0.0085)	0.0129 *** (0.0042)	0.0096 *** (0.0030)
Constant	−3.9187 (2.8633)	−3.6245 * (1.9969)	−0.6064 (0.8975)	0.0357 (0.6133)
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.0969	0.0880	0.0370	0.0354
N	1453	1453	11,007	11,007

Sig. *** p < 0.01, and * p < 0.10. Parentheses contain firm-level clustered standard errors.

. Columns (1) and (2) show the results of the high ESG group: the regression coefficients of analyst logical inconsistency are not statistically significant. Columns (3) and (4) show the results of low ESG group: the coefficients are 0.0129 and 0.0096, both of which are significant at the 1% level. These results show that the positive impact of analyst logical inconsistency on the risk of stock price decline mainly appears in companies with low ESG rating, but has no significant impact in companies with high ESG rating. Companies with good ESG performance usually have very clear governance rules and share a lot of information. These things help them to provide more complete, fairer and newer information for the market, and in this way, the information asymmetry becomes lower. At the same time, higher ESG levels also imply that firms are subject to stricter oversight from investors, regulators, and other stakeholders, constraining management’s space to conceal negative information or engage in opportunistic behavior, thereby weakening the information distortion effects caused by analyst logical inconsistencies.

4.4.4. Types of Logical Inconsistency

The preceding text defines analyst reports with either “optimistic earnings forecast + pessimistic textual interpretation” or “pessimistic earnings forecast + optimistic textual interpretation” as logically inconsistent. Here, we separately examine the impact of these two types of analyst logical inconsistency on stock price crash risk. When $BiasE{F}_{ikt}>0,BiasTon{e}_{ikt}<0$, the analyst report is of the “optimistic forecast + pessimistic tone” inconsistency type, denoted $INCON\_1_{ikt}=1$. When $BiasE{F}_{ikt}<0,BiasTon{e}_{ikt}>0$, it is of the “pessimistic forecast + optimistic tone” inconsistency type, denoted $INCON\_2_{ikt}=1$. The firm-level $INCON\_1_{ikt}$ and $INCON\_2_{ikt}$ are calculated by summing all corresponding reports for the firm. The heterogeneity test results based on the two types of inconsistency are shown in

Table 14. Heterogeneity test results based on two types of inconsistency.

Variables	NCSKEW	NCSKEW	DUVOL	DUVOL
INCON_1	0.0158 *** (0.0058)		0.0124 *** (0.0042)
INCON_2		0.0241 *** (0.0072)		0.0185 *** (0.0052)
Constant	−0.6476 (0.8080)	−0.6888 (0.8094)	−0.0694 (0.5524)	−0.1022 (0.5525)
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.0399	0.0403	0.0367	0.0371
N	12,460	12,460	12,460	12,460

Sig. *** p < 0.01 Parentheses contain firm-level clustered standard errors.

. It can be found that the regression coefficients of both INCON_1 and INCON_2 are significantly positive, indicating that both types of analyst logical inconsistency can increase stock price crash risk. The “optimistic forecast + pessimistic tone” type makes it difficult for investors to accurately judge the firm’s true operating conditions. Some investors may pay more attention to the optimistic forecast while ignoring potential risks, causing negative information to fail to be reflected in stock prices in a timely manner. Conversely, the “pessimistic forecast + optimistic tone” type may weaken market attention to unfavorable performance signals through positive narratives, delaying the release of negative information. Although the two forms differ, both lead to the gradual accumulation of negative information within the firm or in the market. When the accumulation reaches a certain level and is centrally released, it significantly increases future stock price crash risk.

5. Mechanism Analysis

5.1. Financial Risk Channel

Financial risk reflects the stability of a firm’s operating conditions and debt-servicing capacity and serves as an important bridge connecting analyst information and capital market risk. When analyst reports contain logical inconsistencies, investors find it difficult to timely and accurately identify the firm’s true operating conditions, are prone to forming market expectations that deviate from fundamentals, and exacerbate external financing constraints and market volatility, thereby increasing corporate financial risk. The rise in financial risk further weakens market confidence in the firm’s future development, intensifies the accumulation and concentrated release of negative information, and ultimately increases stock price crash risk. Therefore, this paper selects financial risk as a mediating variable to test whether analyst logical inconsistency further affects stock price crash risk by increasing corporate financial risk.

This paper uses the default distance (DD) to measure financial risk. Its economic significance is that the market value of enterprise assets covers the standardized multiple of the nominal value of debt due. This is an inverse indicator. Higher DD means stronger asset reserves to cope with debts, and the possibility of financial crisis due to operational fluctuations is less; that is, the financial risks are lower (Merton, 1974). The basic principle of calculating DD is to subtract the firm’s debt face value from its market value and divide by the estimated firm value volatility. Two assumptions are required: first, that firm value follows a standard geometric Brownian motion; second, that the firm has only one debt liability maturing at time T. Under these two assumptions, DD is calculated through Equations (9) and (10).

$$E=V_{A}N\left(d_1\right)-D{e}^{-rT}N\left(d_2\right)$$

(9)

$${\sigma }_E=\frac{V_A}{E}N\left(d_1\right){\sigma }_A$$

(10)

where $d_1=\frac{ln\left(V_A/D\right)+\left(r+0.5{\sigma }_A^2\right)T}{{\sigma }_A\sqrt{T}}$, $d_2=d_1-{\sigma }_A\sqrt{T}=\frac{ln\left(V_A/D\right)+0.5\left(r-0.5{\sigma }_A^2\right)T}{{\sigma }_A\sqrt{T}}$, E is the market value of equity, $V_A$ is the market value of firm assets, ${\sigma }_A$ is asset volatility, D is the book value of short-term + long-term debt (default point), r is the risk-free rate, and T is the time horizon. $V_A$ and ${\sigma }_A$ are solved iteratively.

This chapter uses the DDmerton indicator and the DDBhsh indicator to measure financial risk. DDmerton is calculated as in Equation (11).

$$DDmerton=\frac{\ln \left(V_A/D\right)+\left(r-0.5{\sigma }_A^2\right)T}{{\sigma }_A\sqrt{T}}$$

(11)

Based on the Merton model, a simplified extended model approximates the asset process using “market leverage + stock return volatility,” thereby avoiding complex implicit asset solution, with the assumption shown in Equation (12).

$$V_A\approx E+D,{\sigma }_A\approx {\sigma }_E·\frac{E}{E+D}$$

(12)

where ${\sigma }_E$ is the stock return volatility estimated by the historical volatility method.

The calculation of the DDBhsh indicator follows the research of Bharath and Shumway (2008), which simplifies the Merton DD and assigns values as shown in Equation (13).

$$DDBhsh=\frac{\ln \left(\frac{E+D}{D}\right)+\left({\mu }_E-0.5{\sigma }_E^2\right)}{{\sigma }_E·\frac{E}{E+D}}$$

(13)

Building on these findings, this study further examines how the propensity for logical discrepancies among analysts can amplify the likelihood of stock market meltdowns via the financial risk conduit. The empirical findings are detailed in

Table 15. Mechanism test results for the financial risk channel.

Variables	DDBhsh	DDmerton	DDBhsh	DDmerton
INCON	−0.1796 *** (0.0195)	−0.1261 *** (0.0198)	−0.1805 *** (0.0186)	−0.1374 *** (0.0184)
Constant	7.7413 *** (0.1358)	6.6185 *** (0.1403)	2.9639 (6.1446)	3.4575 (6.4073)
Controls	No	No	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.1633	0.1235	0.1870	0.1643
N	12,460	12,460	12,460	12,460

Sig. *** p < 0.01 Parentheses contain firm-level clustered standard errors; DDBhsh and DDmerton are measures of financial stability (distance from default); hence, the negative sign indicates the desired direction of impact (increased risk).

. Whether or not control factors are introduced, the coefficients associated with the logical inconsistency of analysts and both DDBhsh and DDmerton demonstrate a robust negative association at the 1% confidence interval. This underscores that a higher degree of analyst inconsistency is a key driver in boosting the risk profile of companies. This result suggests that when there is logical inconsistency between earnings forecasts and textual interpretations in analyst reports, it exacerbates market uncertainty about the firm’s true operating conditions and causes investors to have stronger concerns about the stability of future cash flows, thereby increasing financing constraints and default risk levels, leading to a rise in financial risk. The above results provide important mechanism evidence for the impact of analyst logical inconsistency on stock price crash risk from the financial risk channel, verifying Hypothesis 2.

5.2. Investment Efficiency Channel

Investment efficiency reflects the effectiveness of resource allocation and capital utilization in firms and is an important transmission channel through which analyst information affects firm value. When analyst reports contain logical inconsistencies, they may transmit conflicting information to the market, weakening information transmission efficiency, leading managers and investors to form biased decisions, and thereby inducing overinvestment or underinvestment and reducing investment efficiency. A decline in investment efficiency not only weakens the firm’s future profitability but may also intensify the accumulation of negative information within the firm, making potential operating risks difficult to reflect in stock prices in a timely manner, thereby increasing future stock price crash risk. Therefore, this paper selects investment efficiency as a mediating variable to examine the resource allocation path through which analyst logical inconsistency affects stock price crash risk. In terms of measuring inefficient investment, this paper follows Richardson (2006) and constructs the following model:

$$\begin{array}{l}In{v}_t={\alpha }_0+{\alpha }_{1}Tobin{Q}_{t-1}+{\alpha }_{2}Le{v}_{t-1}+{\alpha }_{3}Cas{h}_{t-1}+{\alpha }_{4}Ag{e}_{t-1}+{\alpha }_5\ln Siz{e}_{t-1}\\ +{\alpha }_6{r}_{t-1}+{\alpha }_{7}In{v}_{t-1}+{\displaystyle \sum Industry}+{\displaystyle \sum Year}+{\epsilon }_t\end{array}$$

(14)

Inv_t stands for the firm’s genuine fresh investment spending in year t, calculated as the sum of total investment minus maintenance investment—essentially, the cash laid out for the acquisition and construction of tangible assets, intangible assets, and other enduring investments, plus the net cash used for buying subsidiaries and other corporate entities. It is offset by any cash received from selling fixed assets, intangible assets, and other long-term investments, minus proceeds from the sale of subsidiaries and other corporate units. Also subtracted are the depreciation of fixed assets, amortization of intangible assets, and amortization of long-term deferred expenses, all divided by the total assets at the start of the year. Industry FE refers to industry-specific fixed effects, with the definitions of other variables remaining consistent with earlier explanations. Equation (14) is estimated on an annual basis through ordinary least squares (OLS) regression. The magnitude of the residual value is a measure of the firm’s inefficient investment levels. A bigger residual figure suggests a more substantial degree of inefficient investment, or, in other words, lower overall investment effectiveness.

Based on the above analysis, this chapter further tests the transmission mechanism of analyst logical inconsistency on stock price crash risk through the investment efficiency channel. In

Table 16. Mechanism test results for investment efficiency and information disclosure channels.

Variables	InefficInvest	InefficInvest	KV	KV
INCON	0.0106 ** (0.0048)	0.0128 *** (0.0048)	0.0121 *** (0.0014)	0.0060 *** (0.0014)
Constant	1.6366 *** (0.0782)	15.8790 *** (1.9675)	0.5443 *** (0.0079)	−3.4092 *** (0.2823)
Controls	No	Yes	No	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
R2	0.2015	0.2353	0.0931	0.1707
N	12,460	12,460	12,460	12,460

Sig. *** p < 0.01, and ** p < 0.05. Parentheses contain firm-level clustered standard errors.

, columns one and two, the regression findings are displayed. Be it with or without control variables in the mix, the regression coefficients reflecting the analyst logical inconsistency in relation to InefficInvest demonstrate a notably positive slope at the 1% confidence level. This implies that there is a notable increase in firm inefficiencies when it comes to investment, thus hampering overall investment efficacy. The logical conflict between earnings forecasts and textual interpretations in analyst reports conveys contradictory information to the market, increases uncertainty in the firm’s external information environment, and decreases capital market information distribution efficiency, thereby enhancing the likelihood of stock market collapses. The above results provide important mechanism evidence for the impact of analyst logical inconsistency on stock price crash risk through the investment efficiency channel, verifying Hypothesis 3.

5.3. Information Disclosure Channel

Information disclosure quality directly determines corporate information transparency and is an important factor affecting capital market resource allocation efficiency and investor decisions. When analyst reports contain logical inconsistencies, they may increase the complexity of the external information environment, reduce the market’s ability to identify the firm’s true information, and weaken the effectiveness of information disclosure, making it difficult for negative information to be timely and adequately transmitted to the capital market. As negative information continues to accumulate, when it is centrally released, it is more likely to trigger sharp stock price declines and crashes. Therefore, this paper selects information disclosure quality as a mediating variable to test whether analyst logical inconsistency promotes stock price crash risk by affecting corporate information transparency. This paper follows O. Kim and Verrecchia (2001) and calculates the KV index using Equation (15). Pt represents the stock price on day t, Vol_t represents the trading volume of the stock on day t, and Vol₀ represents the average trading volume of the stock during the sample period. The coefficient is obtained through OLS regression as the KV value. A larger KV value indicates a stronger correlation between stock returns and changes in trading volume, implying that investment choices are more dependent on trading volume and that the firm’s information disclosure quality is weaker.

$$Ln\left|\left(P_t-P_{t-1}\right){\displaystyle /}{P}_{t-1}\right|=\alpha +\beta \left(Vo{l}_t-Vo{l}_0\right)+\epsilon$$

(15)

In light of the preceding analysis, the study delves deeper into examining how the logic gap among analysts influences the risk of stock market crashes via the avenue of information revelation. Table 16 delineates the regression outcomes, focusing on Columns (3) and (4). Irrespective of whether we include control variables, the regression coefficients for the indicator of logical inconsistency among analysts, when correlated with the measure of information disclosure, yield a significantly positive result at the 1% confidence level. This underscores that logical discrepancies among analysts substantially diminish the quality of corporate information disclosure. The logical conflict between earnings forecasts and textual interpretations in analyst reports increases the complexity of the firm’s external information environment, weakening the market’s ability to accurately judge the firm’s true operating conditions and future development prospects. At the same time, logically inconsistent information also reduces the efficiency of information transmission, making it difficult for external investors to timely obtain and identify the firm’s true value, thereby increasing information opacity and reducing information disclosure quality. The above results provide important mechanism evidence for the impact of analyst logical inconsistency on stock price crash risk through the information disclosure channel, verifying Hypothesis 4.

6. Conclusions

This paper examines whether and how analyst logical inconsistency affects stock price crash risk in China’s A-share market. Using data on 3784 listed non-financial firms and 12,460 firm-year observations from 2016 to 2023, we employ an instruction-fine-tuned large language model to identify the sentiment expressed in analyst reports and construct a measure of analyst logical inconsistency based on the divergence between earnings forecasts and textual tone. The empirical results show that analyst logical inconsistency can increase stock price crash risk. Our analysis utilizes a two-way fixed-effects regression model, revealing a notably positive coefficient for the analyst logical inconsistencies, which is statistically significant at the 1% confidence level. The conclusion remains robust after multiple robustness assessments, including replacing the core independent variable, replacing the dependent variable, considering potential influencing factors, changing the sample period, and considering other fixed effects. At the same time, after using instrumental variable methods to mitigate potential endogeneity issues, the core conclusion remains robust. Second, the impact of analyst logical inconsistency on stock price crash risk varies under different levels of news media attention, corporate reputation, and ESG levels. The promoting effect of analyst logical inconsistency on stock price crash risk is significantly positive under high news media attention, high corporate reputation, and low ESG levels, but insignificant in other firms. In addition, both types of analyst logical inconsistency—“optimistic earnings forecast + pessimistic textual interpretation” and “pessimistic earnings forecast + optimistic textual interpretation”—can increase stock price crash risk. Third, analyst logical inconsistency increases stock price crash risk through three channels: increasing corporate financial risk, reducing corporate investment efficiency, and lowering information disclosure quality.

The findings provide several important theoretical contributions. First, this paper extends the theoretical framework of stock price crash risk by introducing the quality of analyst reports as an important factor in the formation of extreme stock price fluctuations. Existing studies have mainly emphasized firm-level information hoarding, corporate governance, financial reporting quality, investor sentiment, and external monitoring. By contrast, this study demonstrates that the internal coherence of information supplied by financial analysts is also relevant to the accumulation and sudden release of negative information. This finding suggests that stock price crash risk is not determined solely by the information behavior of firms, but also by the quality and structure of information transmission between firms and investors. Second, this paper develops a theoretical transmission framework linking analyst logical inconsistency to stock price crash risk through financial risk, investment efficiency, and information disclosure quality. This framework helps clarify how information conflicts in analyst reports may extend beyond investor interpretation and influence corporate decision-making, market expectations, and the accumulation of firm-specific risks. Third, the results enrich information asymmetry theory by showing that information asymmetry may arise not only from the absence or concealment of information, but also from the inconsistent presentation of information that is publicly available. In this sense, logically inconsistent analyst reports may create a form of informational opacity even when the quantity of disclosed information is relatively high.

This study also contributes to the existing literature in several respects. First, it contributes to the literature on financial analysts by shifting attention from traditional dimensions of analyst performance, such as forecast accuracy, forecast dispersion, recommendation bias, and coverage intensity, to the logical consistency of analyst reports. The findings indicate that the credibility and usefulness of analyst information depend not only on the accuracy of individual forecasts but also on whether quantitative conclusions and qualitative descriptions provide coherent signals. Second, the paper extends the literature on stock price crash risk by identifying analyst logical inconsistency as a new information-intermediary-related determinant. This perspective broadens the understanding of how negative information is accumulated, delayed, distorted, and eventually incorporated into stock prices. Third, the paper contributes to the emerging literature applying large language models to financial research. Compared with conventional dictionary-based methods and earlier sentiment classification techniques, the large language model provides a more flexible approach for recognizing contextual meaning, complex expressions, and the semantic relationship between different components of an analyst report. Accordingly, this study demonstrates how artificial intelligence can be used to transform unstructured textual information into economically meaningful measures of report quality.

The findings also have important practical implications. For regulatory authorities and stock exchanges, greater attention should be paid to the internal consistency and transparency of analyst reports. Relevant institutions may consider improving disclosure standards for earnings forecasts, valuation assumptions, and textual explanations, while establishing more systematic mechanisms for identifying substantial contradictions between quantitative forecasts and qualitative conclusions. The monitoring of analyst reports could be supported by large language models and other natural language processing technologies, thereby improving the efficiency of regulatory screening and reducing the cost of identifying potentially misleading information. For securities firms, internal compliance procedures should strengthen the review of logical consistency, evidentiary support, and the relationship between analysts’ forecasts and narrative expressions. For listed firms, the results suggest that improving financial stability, investment efficiency, and information disclosure quality can help mitigate the adverse consequences associated with inconsistent external information. Investors should also avoid relying mechanically on isolated signals, such as recommendation ratings or emotional wording, and should instead compare earnings forecasts, valuation conclusions, risk disclosures, and supporting explanations in an integrated manner.

Despite these contributions, several limitations should be acknowledged. First, the measurement of analyst logical inconsistency depends on the quality of the training data, sentiment classification rules, model parameters, and threshold selection. Although the results remain robust to alternative thresholds, large language models may still generate classification errors, particularly when analyst reports contain technical terminology, implicit expressions, irony, conditional statements, or mixed sentiment. Second, the sample is limited to Chinese A-share listed non-financial firms during the period from 2016 to 2023. Therefore, the findings may not be directly generalizable to financial firms, private markets, developed capital markets, or other institutional environments.

Future research may extend this study in several directions. First, researchers could construct more comprehensive measures of analyst report consistency by combining multimodal textual analysis, forecast revisions, valuation assumptions, and risk disclosures. Second, future studies could compare different large language models, domain-specific models, and human annotations to evaluate the reliability, interpretability, and cross-model consistency of automated measurements. Third, it would be valuable to investigate whether investor type, analyst reputation, institutional ownership, financial media, and social media moderate the relationship between analyst logical inconsistency and stock price crash risk. Overall, this paper highlights the importance of coherent financial communication and provides evidence that the quality of information interpretation and transmission is an essential component of market stability.