More than words: valuation of words for stock price by using the combination of natural language processing, time-series panel and gradient boosting

Accurately predicting financial market movements remains one of the most challenging tasks in empirical finance and data analytics (Chen and Guestrin, 2016; Hamilton, 2020). While conventional approaches have heavily relied on structured market data and investor-centric sentiment indicators (Brown and Cliff, 2004; Fisher and Statman, 2000), recent advancements in natural language processing (NLP) and machine learning have made it possible to extract market-relevant information from vast quantities of unstructured text data, including news, social media and consumer interactions (Blei et al., 2003; Hutto and Gilbert, 2014). This shift in approach opens new avenues for market prediction models that go beyond price and volume data, incorporating the collective emotions and complaints of market participants and consumers.

In addition to the NLP, sentiment analysis can be merged into machine learning. The traditional sentiment studies have primarily focused on signals derived from institutional or individual investor behavior, such as survey-based indices, mutual fund flows and trading patterns (e.g. Lee et al., 1991; Schmeling, 2009). However, emerging evidence suggests that consumer complaints and emotional expressions are expected to capture central aspects of market mood. The capture of market mood often acts as an early warning signal for firm risk, reputation crises and abrupt changes in stock price dynamics (Gruca and Rego, 2005; Luo and Homburg, 2008). A recent study shows that reputation indicators constructed from large volumes of online customer comments are statistically associated with bank stock price movements, reinforcing the relevance of text-based reputational signals in financial markets (Casado-Molina et al., 2025). Based on the expectation, recent developments in machine learning and ensemble modeling have further enhanced the ability to integrate such qualitative information into quantitative forecasting frameworks systematically (Zięba et al., 2016). Despite these advancements, limited research combines NLP-derived consumer complaint and emotion data with advanced time-series panel models and machine learning techniques to forecast stock price movements. Existing studies in behavioral finance and market sentiment have demonstrated that emotions and collective mood can impact market volatility, risk-taking and valuation anomalies (e.g. Baker et al., 2012; Barber and Odean, 2008). However, most have not systematically quantified the predictive power of consumer-generated textual signals or explored their integration with financial and econometric indicators (e.g. Gouthier and Schmid, 2003; Grönroos, 1984).

To address this gap, the present study develops a comprehensive modeling framework that harnesses NLP techniques such as Latent Dirichlet Allocation (LDA) and VADER sentiment analysis (Blei et al., 2003; Hutto and Gilbert, 2014). LDA extracts latent topics and sentiment analysis extracts emotional tone from large-scale consumer complaint data. These textual variables can be operationalized on a quantified scale, which can be integrated into time-series panel models (Hamilton, 2020). These models are eventually combined with ensemble-based machine learning approaches (Chen and Guestrin, 2016). To summarize, it is possible to construct hybrid models that capture nonlinear and complex relationships among sentiment, complaints and stock market outcomes.

This research has three primary objectives: (1) empirically evaluate the predictive power of consumer complaints and emotions in forecasting stock price movements; (2) demonstrate the value of integrating unstructured text data with traditional financial indicators in a robust analytic framework and (3) offer actionable implications for investors, risk managers and corporate decision-makers seeking to leverage real-time sentiment signals for enhanced market forecasting and reputation management (Sirignano and Cont, 2021; Tetlock, 2007). This study contributes to the academic literature and the evolution of predictive analytics in financial markets by bridging the gap between behavioral finance theory and practical data science.

A central theme of behavioral finance is the recognition that market participants frequently rely on intuitive and emotion-driven judgments (Loewenstein et al., 2001; Slovic et al., 2002). Among various concepts within the field of behavioral finance, affect heuristics describe how individuals use their immediate emotional responses as cognitive shortcuts to evaluate risks, benefits and probabilities. Investors and stakeholders may interpret information about firms or markets based on factual content in financial contexts. The interpretation may also involve the emotional tone accompanying such information.

This perspective is one of the baseline concepts of the present study. Specifically, consumer complaints and emotional expressions represent direct manifestations of affective reactions to market events and firm behaviors. Prior psychology and behavioral economics research suggests that such emotional signals can influence individual decision-making and aggregate market trends. The decision-making and the aggregation of trends are more likely to act as catalysts for overreaction, underreaction or herding behavior (Loewenstein et al., 2001). By systematically quantifying these affective signals from large-scale text data, the present research seeks to capture dimensions of investor psychology that are often invisible to conventional models.

Investor sentiment theory argues that the overall mood of market participants can cause asset prices to deviate from their fundamental values (Baker et al., 2012; Brown and Cliff, 2004). The traditional approach, such as efficient market hypothesis, explains that markets efficiently and rationally incorporate all available information. However, sentiment theory posits that investor sentiment can systematically bias market outcomes through waves of optimism and pessimism and a collective mood. Sentiment-driven mispricing, volatility clustering and periods of irrational exuberance or fear are sometimes utilized as market indicators instead of rationality (e.g. Baker et al., 2012).

The present study extends the range of sentiment proxies within this framework by focusing on consumer complaints and emotional tone in textual data. While prior studies have relied on indicators such as survey data, news articles or investor flows (Fisher and Statman, 2000; Schmeling, 2009), this research posits that consumer-generated expressions can be operationalized as an informative signal of underlying market sentiment. In this way, affective signals captured via NLP become directly operationalized as sentiment variables, connecting the micro-level psychology of individuals to macro-level market phenomena.

A further refinement within behavioral finance is the theory of limited attention. The theory explains that market participants possess bounded cognitive resources. Because of these bounded cognitive resources, investors disproportionately attend to salient, emotionally charged or widely disseminated information (Barber and Odean, 2008; Hirshleifer and Teoh, 2003). In a media-saturated environment, emotionally intense signals, such as widespread consumer complaints or sharp shifts in emotional tone, are more likely to capture collective attention and prompt rapid behavioral responses.

This theoretical insight provides a critical link between the raw presence of affective signals and their actual impact on market dynamics. In other words, no market equally processes all information. Instead, information (e.g. viral complaints or trending consumer sentiment) can disproportionately influence trading volume, volatility and price adjustments. By extracting and analyzing these signals in real time, the present study leverages the salience and urgency of consumer-based affect as a mechanism for predictive modeling.

Adopting NLP for financial market forecasting has grown rapidly in recent years. Early research explored how textual information (i.e. financial news, analyst commentary and regulatory filings) could help extract signals anticipating future price movements (Loughran and McDonald, 2011; Tetlock, 2007). These studies showed that the polarity and tone embedded in financial texts could precede abnormal returns and shifts in trading volume. The expansion of social media and online communities has further broadened the field. Digital platforms (e.g. Twitter, StockTwits and various investor forums) were considered real-time barometers for market sentiment (Bollen et al., 2011; Sprenger et al., 2014). For instance, Bollen et al. (2011) proved that aggregate moods from social media content predicted moves in the Dow Jones Industrial Average. Similar works (e.g. Chen et al., 2014; Oliveira et al., 2017) found that the added value of sentiment indicators derived from online investor discussions can improve the explanatory power of standard market models.

More recently, methodological developments in NLP have allowed researchers to mine increasingly complex relationships from financial texts. LDA and sentiment algorithms like VADER were utilized to classify topics (Hutto and Gilbert, 2014). In the meantime, adopting deep learning architectures has enabled the capture of nuanced context and sequential patterns in financial narratives (e.g. Ashtiani and Raahemi, 2023; Ding et al., 2015). Empirical results indicate that blending these text-based features with conventional financial variables frequently yields more accurate forecasts of both returns and volatility (Hu et al., 2018; Si et al., 2013).In the banking context, text-mining research further demonstrates that unstructured online customer reviews can be decomposed into distinct service attributes and sentiment dimensions, thus enabling systematic extraction of evaluative information beyond anecdotal complaints (Mittal and Agrawal, 2022).

Despite these advances, several challenges persist. Interpreting the mechanisms behind NLP-driven forecasts and ensuring their effectiveness in real-time trading environments remains complex. Integrating various text sources (e.g. news, social media and consumer complaints) while accounting for factors like time lags and information diffusion continues to pose open questions for practitioners and scholars (Sirignano and Cont, 2019). The accumulating evidence points to NLP as a valuable supplement to traditional analytics, equipping risk managers and traders with tools for more responsive decision-making.

Within economic sociology, emotions are no longer considered incidental to economic life. Contemporary scholarship suggests that affective responses are central to calculation, bargaining and decision-making and that they are inseparable from cognitive processes. Organizations exhibit rational and emotional characteristics that reflect their members' aggregated outlooks and expectations.

The concept of investor sentiment has become a fixture in analyzing stock market dynamics. A significant body of work has investigated the antecedents of sentiment and its empirical influence on asset prices (Fisher and Statman, 2000; Gao et al., 2024). These analyses consistently demonstrate that sentiment indices can forecast returns and serve practical forecasting purposes. It means that constructing sentiment proxies has been shown to predict shifts in stock prices (Bollen et al., 2011). In the meantime, other research has emphasized the role of lagged sentiment in market forecasting (Chung et al., 2012). Findings from diverse research streams reinforce the view that changes in investors' psychological and emotional states are key drivers of market behavior. Both immediate and lagged investor mood measures reliably signal future return shifts (e.g. Bollen et al., 2011; Chung et al., 2012; Fisher and Statman, 2000; Gao et al., 2024). Consistent with this perspective, prior research shows that inventors' stock performances are shaped by expected financial returns as well as by subjective evaluations of firms' products and brands, which indicates that nonfinancial perspectives play an important role in stock-related judgements (Aspara and Tikkanen, 2010). Recognizing that institutional and international sentiment and retail investor mood play a formative role in market outcomes has led to a more comprehensive understanding of market sentiment as a collective phenomenon.

Despite accumulating evidence supporting the predictive power of investor sentiment, much prior empirical research has focused on a single facet, such as sentiment among individual investors. This narrow focus has limited the broader applicability of findings, as market returns are influenced by a constellation of sentiments originating from individuals, institutions and foreign market participants. More recent scholarship has adopted a multi-dimensional approach by incorporating sentiment indicators from all major investor categories to more comprehensively forecast stock index returns.

Beyond investor sentiment, accurate measurement of the effects of marketing strategies on a firm's market value remains a central objective within business research. The evolution of financial metrics in this context is well documented with scholars increasingly leveraging financial indicators to quantify the impact of managerial actions on firm value (Mintz and Currim, 2013; Srinivasan and Hanssens, 2009). Other research supports the holistic evaluation of firm performance (Luo and Homburg, 2008; Osinga et al., 2011). Luo and Homburg (2008) observe that the pursuit of perfect customer satisfaction is rarely achievable. Increased dissatisfaction does not always translate into an immediate decline in market value. However, persistent neglect in tracking the financial consequences of marketing activities may erode firm credibility and long-term value (Mintz and Currim, 2013; Rust et al., 2004). Consequently, managing customer complaints and satisfaction effectively is essential for risk mitigation and strategic value creation. Likewise, empirical findings consistently illustrate a positive association between customer satisfaction and shareholder value, reinforcing the conceptual link between market value and the quality of customer relationships (e.g. Gruca and Rego, 2005; Luo and Homburg, 2008). Several studies have sought to establish direct evidence that marketing initiatives can affect a firm's financial statements and that customer retention and satisfaction are integral to realizing future value. High customer satisfaction increases retention rates, reducing future transaction costs and supporting sustainable growth.

Individuals' and groups' emotional and rational judgments fundamentally shape a company's value and public image. Organizational decision-making is influenced by the emotional climate within and around the company. In the bank marketing domain, banking reputation is increasingly conceptualized as a multidimensional and strategically consequential construct shaped by stakeholder perceptions and trust, reinforcing its role as a salient nonfinancial signal (Carè et al., 2024). Such reputational perceptions are not merely abstract constructs but are increasingly observable through customer-generated textual data that systematically capture customer evaluations and sentiment patterns (Ghadiridehkordi et al., 2025). Accordingly, consumers' dissatisfaction has far-reaching consequences for brand reputation and market value (Grönroos, 1984). Exposure to negative word-of-mouth can produce cognitive biases among potential customers, with those encountering negative information developing more adverse perceptions than those exposed to neutral or no information (De Matos and Rossi, 2008). Companies that fail to meet customer expectations see reduced ability to acquire and retain clientele, which ultimately impedes future cash flow. Incremental improvements in customer retention can lead to substantial gains in market value; for example, a 1% increase in retention may boost firm value by 3%–7% (Gupta et al., 2004). These findings collectively affirm that the customer base and its satisfaction level are intangible assets that significantly shape a company's prospects (Rust et al., 2004).

The existing literature firmly establishes that rational and emotional drivers influence stock prices and firm valuations substantially. Integrating psychological and behavioral economics perspectives into empirical research has enriched the understanding of how market dynamics are shaped by sentiment, satisfaction and collective emotion.

Many financial firms are beginning to use AI-based tools in their operations. They may use AI-based tools for various purposes (e.g. trading and risk analysis, credit assessments and fraud monitoring) while moving away from conventional rule-based frameworks toward systems. Primarily, the rule-based framework was driven by large-scale data and statistical learning. Forecasting corporate financial distress has become a significant focus, drawing the attention of scholars and industry practitioners (Alaka et al., 2018; Altman et al., 2017).

A wide array of computational methods has emerged to meet this challenge. For example, artificial neural networks (ANNs) have been valued for capturing nonlinearities and subtle interactions in financial variables (Atiya, 2001; West, 2000). With their solid theoretical basis, support vector machines (SVMs) have also proved effective in bankruptcy prediction and risk modeling (Sulistiani et al., 2022). Techniques such as decision trees (DTs), random forests (RFs) and their ensemble combinations have improved model accuracy and robustness by blending the predictions of many base learners (Barboza et al., 2017; Chou, 2019; He and Fan, 2021; Sun and Li, 2008). Progress in deep learning has brought additional gains. Research has shown that deep neural networks can surpass traditional linear models in predicting asset prices, particularly in markets where complexity and nonlinearity are pronounced (Sirignano and Cont, 2019). In high-frequency trading, AI models have been used for split-second, automated decisions, often reacting to new data far more quickly than human traders (Nguyen et al., 2015). This data-driven approach has made it easier for institutions to evaluate financial soundness.

Additionally, gradient boosting decision trees have shown exceptional performance in forecasting financial time series and structured data (Amirshahi and Lahmiri, 2024; Zięba et al., 2016). Ensemble algorithms, including gradient boosting, have become increasingly popular due to their ability to balance bias and variance in complex datasets. Gradient boosting and its varieties approach consistently outperform more basic machine learning techniques on both speed and predictive accuracy (Chen and Guestrin, 2016; Mitra et al., 2022). These methods effectively address underfitting and overfitting and can integrate market data and financial indicators. Even with behavioral factors, these models perform well and are indispensable in today's financial analytics landscape. The evolution of AI and machine learning continues to drive advances in financial prediction and risk management, with gradient boosting algorithms setting the benchmark for model performance and adaptability in dynamic, data-rich environments.

This study utilized NLP for two key aspects. The first is the rate of consumer complaints, and the second is the emotional intensity of these complaints. Traditional market performance data did not adequately reflect these metrics. However, with advances in computer science and algorithm development, it is now possible to quantify textual data using NLP. This study employs two NLP methodologies to measure the concentration and directionality of opinions.

LDA is used to quantify the concentration of financial consumers' complaints. LDA is an introduced topic modeling method that identifies latent information within a text corpus. It is a generative, probabilistic, Bayesian, unsupervised machine learning algorithm that uses a Dirichlet distribution to assign words in documents to topics (Grün and Hornik, 2011). LDA analysis enables the exploration of hidden topics in social data, supporting a multifaceted research approach. By identifying associated keywords through keyword network analysis, LDA expands the scope of research to include various applications (Blei et al., 2003).

The LDA process extracts meaningful topic groups with similar keywords from the given text dataset (Jelodar et al., 2019). From complex datasets, each keyword in the text corpus is distributed among multiple topic groups (⁠$zi$⁠), where the number of topic groups can be various (⁠$i≥1)$⁠. Researchers input different possible values (i) for $zi$ and select the best performance (i.e. optimization). After optimization, the algorithm reassesses each word's topic assignment based on the overall word distribution. The process continues until a stable pattern emerges, in which words no longer shift between topics.

During LDA, the marginal topic probability (MTP) can be estimated as one of the keyword-group indicators. MTP denotes the extent to which each keyword group dominates the given documents. In other words, MTP represents the probability of each keyword group appearing within a document by using the following Equation (1):

where P is the marginal probability of a keyword group found by MTP; i is the ith keyword group; M is the total document given as a text dataset; m is the individual document file.

Therefore, when multiple documents are analyzed together within the same period, several keyword groups (⁠$zi$⁠) are identified by considering all the documents (M). Among these groups, the keyword group with the highest proportion in a document can be identified, and this study will use the probability of this keyword group. The keyword group with the highest proportion indicates the extent to which people's complaints are connected to the same issue. In other words, if the proportion of the highest-ranking keyword group is high, many complaints are about the same problem. Conversely, if the proportion of the highest-ranking keyword group is low, it indicates that the complaints during that period are not converging on a single issue. Therefore, MTP is transformed and used as an indicator, like the Equation below (2):

Sentiment analysis quantifies the extent to which consumers express negative emotions. Even if there are multiple sentiment analyses within NLP (Khan et al., 2010), this study uses the Valence Aware Dictionary for Sentiment Reasoning (VADER) sentiment classifier to label data with negative and positive sentiments (Hutto and Gilbert, 2014). The sentiment score is calculated as a degree representing positive and negative sentiment by following Equations (3) and (4):

in the condition of ${S(wi)>0:positivewordS(wi)<0:negativeword$

where N is the total number of words in the sentence for positive and negative; $S(wi)$ is the sentiment score of positive and negative word (⁠$wi$⁠). VADER generates sentence scores based on the sentiment of words within the sentence to give a systematic understanding of unstructured text. Through sentiment analysis, underlying sentiments in customer complaints can be revealed to help provide a structured method for understanding opinions hidden within the text.

Logarithmic transformation is widely used in this research for monetary variables, including stock price, assets, debt, equity, EBIT and revenue. Logarithmic transformation helps to address several econometric issues: (a) it reduces the influence of extreme values and skewness commonly found in financial data, thereby mitigating heteroscedasticity and improving the validity of statistical inference (Wooldridge, 2016); (b) it allows for the interpretation of regression coefficients as elasticities or approximate percentage changes, which enhances economic interpretability and comparability across firms of different sizes (Gujarati and Porter, 2009); (c) it normalizes distributions, making them closer to the assumptions underlying linear regression and other statistical techniques. The logarithmic form has become standard in empirical studies of company valuation, capital structure and corporate performance (Booth et al., 2001; Fama and French, 2002). The logarithm was made by using the following Equation (5):

Time-series panel path modeling provides a robust framework for examining how current financial outcomes, such as stock price, are dynamically influenced by past values of both endogenous and exogenous variables across multiple companies and over time (Hamilton, 2020). This approach has been widely adopted in financial market research to capture the autoregressive and cross-sectional dependencies in panel data. The approach is beneficial for disentangling the temporal effects of investor sentiment and behavioral variables (Brown and Cliff, 2004; Lee et al., 1991; Schmeling, 2009).

A core motivation for adopting this approach is the well-documented empirical finding that stock prices, investor sentiment and firm-specific variables (e.g. asset, debt, equity, EBIT, revenue) exhibit substantial autoregressive features. These autoregressive processes account for the persistence of economic shocks, information diffusion and behavioral feedback mechanisms that drive individual and aggregate market outcomes (Tetlock, 2007). Furthermore, as investor sentiment and emotion can influence and be influenced by lagged price dynamics and news events, it is crucial to incorporate these time-lagged structures in any empirical modeling (Naseem et al., 2021). Recent methodological advances in panel structural equation modeling (SEM), latent curve modeling and dynamic SEM (Bollen and Curran, 2006; McArdle, 1988) have enabled researchers to estimate these autoregressive relationships and contemporaneous paths in a unified framework. Latent curve and dynamic SEM models, as well as recent developments in dynamic structural equation modeling for panel data, allow for the joint estimation of autoregressive features in both outcomes (e.g. stock price) and exogenous predictors (e.g. complaints and emotions), including the possibility of correlated measurement errors and unit-level random effects. In this study, the stock price is assumed to be associated with its own lagged value and with the lagged intensity of complaints and emotions, as captured by MTP and sentiment scores from previous periods. This structure reflects both the behavioral feedback loop emphasized in the investor sentiment literature and the persistence of shocks identified in time-series econometrics.

Since stock price is highly correlated with each company's valuation indicators (i.e. asset, debt, equity, EBIT and revenue), these are also included and controlled as covariates (Booth et al., 2001). The general structure of the time-series panel path modeling equations used in this study is as follows Equation (6):

where i is each company from the panel; t is the time; $yit$ is the observed outcome variable (logarithm format of price) for each company(i) at time t; $ηit$ is the company(i)'s valuation indicators at time t; $Ny$ is lagged value (t-1) of price for autoregressive; $δi(t−1)$ is lagged value (t-1) of complaints and emotions; $Λy$ and $My$ are the factor loadings of $ηit$ and $δi(t−1)$ respectively; $ϵit$ is the measurement error.

Complaints and emotions are modeled as exogenous panel time-series factors, each displaying strong autoregressive features (Bollen and Curran, 2006; McArdle, 1988). Their dynamics are captured in Equation (7)

where $B$ is the coefficient matrix of complaints and emotion; ξ is the structural disturbance. Complaints are represented by the highest and lowest marginal topic probabilities, while emotions are quantified as positive and negative sentiment scores. The log format is used for monetary variables to reduce heteroskedasticity and skewness, and to allow coefficients to be interpreted as elasticities.

Applying the above logic, the empirical models in this study estimate the predictive effects of lagged complaints and emotions, as well as firm valuation indicators, on stock price dynamics. The key valuation indicators include assets, debt, equity, EBIT and revenue. The complaint variables are captured using the highest and lowest marginal topic probabilities, whereas positive and negative sentiment scores represent emotional valence. Therefore, the primary regression specification for the price equation takes the following form, Equation (8):

where j indicates the complaints and the emotions, including MTPH, negative sentiment and positive sentiment; v is the company's valuation indicators, including asset, debt, equity, EBIT and revenue; c is the company's control factors, including state and industry; Month is the time; $ϵit$ is the error term. Regarding the lagged values of complaints and emotions, they are exogenous factors that were not unit-identified (i). As the basic Equation (6) changed to an empirical model of Equation (8), the basic format of Equation (7) becomes the empirical Equations (9) to (11):

where mtph and h are the highest MTP score from the groups found; neg and n are the negative sentiment score; pos and p are the positive sentiment score; Month is the time factor; $ξh(t−1)$⁠, $ξn(t−1),andξp(t−1)$ are error terms at time (t-1).

This structure allows for the dynamic interdependence among complaints, emotions and stock prices. It aligns with the panel data literature, which shows that both economic and behavioral variables exhibit significant temporal persistence and cross-lagged effects. By employing a panel path modeling strategy, this study accounts for both autoregressive price dynamics and exogenous shocks from the complaint and emotion measures, thereby enhancing the explanatory and predictive power of the empirical framework.

The Time-Series Panel Path Modeling constructed in the previous section can be seamlessly integrated into a Gradient Boosting framework (Friedman, 2002; Natekin and Knoll, 2013, , 2013). To incorporate the time-series path modeling into Gradient Boosting, the predictors from Equations (8) to (11) are clarified as follows, as in Equation (12):

Among the list from $xt$⁠, the list within the bracket are utilized as estimated values like $δˆ(t−1),mtph,δˆ(t−1),neg,andδˆ(t−1),pos$⁠. The factors of $xt$ that have converged and are used as predictors in the fundamental equation of Gradient Boosting. The fundamental functions of Gradient Boosting (Chen and Guestrin, 2016; Friedman, 2002) are as follows, Equations (13)–(16).

where $rj,m$ is the residual of mth iteration for jth data point; $yj$ is observed outcome for jth data point; $Fm−1(xj)$ is predicted outcome for jth data point when the model is (m-1)th iteration; $L(yj,Fm−1(xj))$ is loss function for the difference between mth observed y and the (m-1)th predicted y (Natekin and Knoll, 2013). To minimize the residuals, each weak learner $hm(x)$ is fit to the negative gradient (the pseudo-residuals) using Equation (14):

where $hm(x)$ is the learning model with m-th iteration; argmin is the argument for minimum. Therefore, the Equation minimize the residuals (⁠$rj,m$⁠) from the weak model (⁠$h(xj))$ to the iterated m-th model. The iterated model can update the prediction model like below Equation (15) and the prediction repeats as shown in Equation (16):

where $Fm−1(x)$ and $Fm(x)are$ the prediction models when it is (m-1)th and mth iteration, respectively; v is learning rate; M is the total iteration number. Finally, the factor list (⁠$xt$⁠) utilized within Equation (16) like Equation (17):

This approach allows for the integration of dynamic, path-dependent features (such as lagged complaints and emotions) with fundamental and control variables, leveraging the powerful predictive capabilities and flexibility of Gradient Boosting (Chen and Guestrin, 2016; Friedman, 2002). Notably, GB can flexibly accommodate nonlinearity, interaction effects and the importance of heterogeneous features. GB is particularly suitable for complex, high-dimensional financial prediction tasks (Natekin and Knoll, 2013).

As shown in Figure 1., the above explanation for empirical modeling was made for Model D as the full model. However, conducting only Model D can't confirm how the complaints significantly contribute to the findings. Therefore, three models (i.e. Model A, Model B, and Model C) are added to be compared. Specifically, in Stage 1, two models are compared. Model A is conducted without complaints and emotions, but Model B is conducted with complaints and emotions. The findings from the comparison between Model A and Model B are expected to show the significant meaning of adding complaints and emotions. Therefore, the first hypothesis is as follows:

The Time-Series Panel Analysis is used for the first hypothesis by using Equation (8) for Model B and Equation (18) for Model A:

After the Time-Series Panel comparison stage (i.e. Stage 1 in Figure 1) is completed, a prediction comparison is conducted in Stage 2, as shown in Figure 1. In the prediction comparison, the first prediction was conducted using the Time-Series Panel Analysis for Models A and B (See H2 in Figure 1). The hypothesis for this prediction is:

The second prediction was conducted using a combination of Time-Series Panel Analysis and Gradient Boosting, with a different set of predictors (See H3 in Stage 3, Figure 1). For the full model (i.e. Model D), the combined Gradient Boosting with Time-Series Panel Analysis (See Equation (17)) using the full predictor list (See Equation (13)) is employed. To compare the prediction effect of complaints and emotions, Model C with the combined Gradient Boosting with Time-Series Panel Analysis uses the reduced predictor list as shown in Equation (19):

The last hypothesis is:

By comparing Model B and Model D (see H4 at the bottom of Figure 1), the combined prediction rate (i.e. Time-Series Panel Analysis and Gradient Boosting) can be compared with the simple prediction using only Time-Series Panel Analysis. Therefore, the last hypothesis is:

The prediction performance check is tested by using RMSE, MAE and MAPE, which are explained in the next Methodology section.

This study utilized two integrated data sources covering August 2018 to July 2023. First, consumer complaint narratives were obtained from the Consumer Financial Protection Bureau (CFPB), which aggregates and anonymizes public reports of issues with financial products and services. After excluding records lacking text, more than 1 million complaint documents were analyzed using advanced NLP techniques, with a focus on topic modeling and sentiment extraction. This process yielded monthly indicators of complaint topic intensity and associated sentiment scores across the financial sector.

The second dataset comprises firm-level financial and market information collected from SandP Capital IQ. This database provides detailed monthly data over the same 60-month window for publicly traded financial firms, including banks, insurance companies, capital market institutions, consumer finance entities and mortgage real estate investment trusts. Only companies with complete reporting on key financial variables, such as equity, assets, debt, EBIT, revenue and capital size, were retained for analysis. The final sample comprises 261 finance-related firms, each observed for up to 60 months, yielding a balanced panel structure.

Monthly aggregates of consumer sentiment, as well as topic probabilities for each dominant and weak complaint theme, were merged with firm-level panel data with the corresponding month used as the time index. This integration enabled the construction of a comprehensive dataset linking shifts in collective investor mood, as reflected in consumer complaints, to financial fundamentals and stock price dynamics of major publicly listed financial companies.

All variables related to monetary values, such as stock prices, capital, liabilities, equity, EBIT, and revenue, were transformed using natural logarithms to address potential skewness and heteroskedasticity and to facilitate elasticity-based interpretation of regression coefficients. The final panel data allow for empirical modeling of the relationship between complaint-derived mood metrics and firm valuation outcomes at a high temporal resolution, supporting robust time-series and panel data analyses.

However, it is important to acknowledge the limitations of this analysis. The complaint-based indicators (MTPH and sentiment) are constructed at the monthly market level by aggregating CFPB complaint narratives across the financial sector and then merged to firm-month observations using the time index. Accordingly, the inferences in this article concern temporal associations between market-level consumer discourse and firm-level stock prices rather than complaint-level causal effects or firm-specific recovery processes.

From August 2018 to July 2023, more than 1.16 million text-based complaints were submitted to the Consumer Financial Protection Bureau (CFPB) regarding financial products and services. The most frequent complaint categories throughout the observation period included credit reporting, general-purpose credit cards and checking accounts. Complaint volume increased substantially over time, with annual totals rising from approximately 84,000 in late 2018 to over 740,000 in 2022. As shown in Table 1, each month, LDA identified between four and six distinct complaint topics, with the dominant topic's marginal topic probability (MTP) ranging from approximately 0.25 to 0.41, indicating substantial clustering of consumer issues. Sentiment analysis further revealed negative and positive sentiment scores, which were used as instrumental variables in subsequent analyses.

The firm-level sample comprises 261 publicly traded financial companies observed monthly over five years. Table 2 presents summary statistics for key variables after applying a natural-log transformation. The mean logged daily stock price was 3.42 (SD = 1.33), with a minimum of 0 and a maximum of 9.48. The average logged capital was 7.20 (SD = 1.75), while the mean logged equity and liabilities were 6.97 (SD = 1.85) and 4.55 (SD = 1.98), respectively. Logged EBIT and revenue also exhibited broad ranges, reflecting considerable variation in firm size and performance. The sample covers various financial subsectors, including banking, insurance, consumer finance and mortgage REITs, and it is geographically diverse across multiple U.S. states.

For each month, the analysis identified both the number of distinct complaint topics and the intensity of topic clustering, as measured by MTP. For example, in July 2023, four dominant topics were observed, with the highest MTP (0.411) indicating that over 41% of complaints belonged to the most common topic cluster. Sentiment analysis for the same month yielded negative-to-positive sentiment ratios of 0.085 and 0.092, respectively.

Based on the empirical modeling above, hypothesis testing is organized around a series of progressively sophisticated analytic approaches with each model tailored to the corresponding research question. All analyses utilize the merged monthly panel dataset with the primary dependent variable defined as the natural logarithm of each firm's average monthly stock price. The main independent variables include the MTP for the dominant/weakest complaint clusters from LDA and sentiment scores (positive and negative) derived from sentiment analysis. Firm-level controls, such as log-transformed assets, debt, equity, EBIT and revenue, are incorporated, along with categorical indicators for industry, state and calendar month, to account for unobserved heterogeneity.

Two complementary modeling strategies are employed to address Hypothesis 1 (H1). Model A employs a fixed-effects panel regression framework to estimate contemporaneous and lagged associations among complaint intensity, sentiment and stock price, controlling for firm and temporal fixed effects. Model B adopts a structural decomposition strategy (panel path modeling) to disentangle dynamic relationships among complaints, emotions and stock prices, explicitly modeling lag structures and potential mediation effects.

The focus shifts to predictive validation for Hypotheses 2 and 3 (H2, H3). The dataset is divided into equally sized training and test subsets with model parameters estimated on the training set and predictive accuracy assessed on the holdout sample. This approach enables an out-of-sample evaluation of each model's generalizability. In the case of H3, Gradient Boosting (GB) algorithms are employed with model parameters (e.g. learning rate, number of trees and maximum depth) tuned via cross-validation to optimize predictive performance.

Finally, Hypothesis 4 (H4) is addressed through robustness and sensitivity analyses, drawing on the core model specifications already described. No additional estimation procedures are required for this hypothesis; the goal is to ensure the stability and reliability of the principal findings across alternative specifications and subsamples.

To assess predictive accuracy, this study employs three widely used error metrics: root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), as defined in Equations (20) and (22). These metrics offer complementary perspectives on prediction error magnitude and distribution. RMSE penalizes larger errors more heavily by squaring the differences between predicted and actual values, whereas MAE captures the average magnitude of errors in absolute terms. MAPE expresses the prediction error as a percentage of the observed value, which facilitates interpretability across contexts. These criteria are particularly recommended for continuous dependent variables and are standard practice in forecasting literature when comparing predictive model performance (Hyndman and Koehler, 2006; Willmott and Matsuura, 2005). Reporting RMSE, MAE, and MAPE provides a comprehensive view of both absolute and relative prediction accuracy.

Table 2 reports the results from two complementary models that assess the relationship between lagged price, complaint concentration, sentiment and stock price dynamics. Model A demonstrates that the previous month's stock price and EBIT are positively associated with the current price. It suggests the significance of profitability in explaining stock price variation. In contrast, revenue is negatively associated with stock price, potentially indicating that revenue increases, absent corresponding increases in profitability, do not necessarily translate into higher market valuation for financial firms.

Model B further explores the dynamic influence of complaints and emotions (see Table 2). The marginal topic probability for the dominant complaint cluster (MTPH) is positively and significantly associated with current price, suggesting that a greater concentration of complaints on a single topic may reflect heightened market concern or attention. This, in turn, is linked to price increases. Notably, although negative sentiment is not significantly associated with price, positive sentiment extracted from complaints is positively and significantly associated with price. This suggests that, even in the context of complaints, positive signals can be interpreted by the market as indicators of a favorable outlook, thereby supporting price appreciation.

In addition, Model B reveals intricate temporal relationships among complaints and emotional signals. Specifically, a high concentration level in the previous period (past MTP) predicts increased current concentration and heightened negative sentiment. This pattern implies that adverse attention may intensify and persist over time when complaints are clustered. Conversely, prior high concentration reduces positive sentiment, reinforcing that collective concern may crowd out positive discourse. Moreover, negative sentiment appears to diminish overall concentration, which indicates that the resolution or expression of negative emotions helps diffuse group focus over time. Additionally, negative sentiment increases subsequent negative and positive emotions, suggesting a complex emotional contagion effect. Finally, positive sentiment increases current concentration and elevates subsequent positive sentiment while reducing negative sentiment, and this highlights the mutually reinforcing nature of emotional dynamics within complaint discourse.

Collectively, these results underscore the significance of the affective and cognitive dimensions of consumer complaints in shaping stock price movements and reveal a nuanced interplay among emotion, attention and market valuation in the financial services sector.

The structural decomposition in Table 3 reveals several key pathways through which past complaint concentration and sentiment dynamics influence current stock prices. First, a higher marginal topic probability for dominant complaints two periods prior (MTPH_t−2​) exerts a significant positive effect on current price, directly and indirectly. Specifically, MTPH_t−2​ increases MTPH in the prior period, elevating current prices, highlighting the persistence and amplifying effect of collective attention within complaints on market valuation. Second, negative sentiment from two periods prior (Negative_t−2​) exhibits a complex pattern of influence. While it is associated with a reduction in current price through its adverse effect on MTPH, it also contributes to increased positive sentiment in the intervening period, subsequently raising prices. This dual pathway suggests that expressions of negative emotion may, over time, stimulate constructive discourse or managerial response, ultimately yielding positive valuation effects. Third, positive sentiment in the prior periods (Positive_t−2​) shows direct and indirect pathways to higher current prices. Directly, Positive_t−2​ increases current price, and indirectly, it amplifies both prior MTPH and positive sentiment, each of which also positively affects current price. This underscores the enduring and compounding impact of sustained positive emotion in complaints on market outcomes. These findings demonstrate the intricate and temporally layered ways in which complaint concentration and sentiment reflect and shape financial market performance over time.

Table 4 presents a comparative performance of multiple gradient-boosting model specifications on the training and test (validation) datasets. Evaluation criteria include RMSE, MAE and MAPE, facilitating a comprehensive assessment of absolute and relative prediction accuracy.

Across the tested configurations, the “Extreme” gradient-boosting implementation consistently achieves the lowest prediction errors on both training and test data. For example, the RMSE for the Extreme model is 0.013 on the training set and 0.065 on the validation set, outperforming the standard Scikit-learn implementation and alternative approaches such as “Extreme with RF” and “Category.” MAE and MAPE results similarly indicate superior accuracy for the Extreme model, with test MAE values as low as 0.039 and MAPE values of 0.091, highlighting the model's robustness and generalizability.

In contrast, models combining extreme boosting with random forest (Extreme with RF) exhibit markedly higher error rates (e.g. test RMSE ≈ 0.80 and MAPE >1.28), indicating potential issues with overfitting or incompatibility for this specific financial time-series prediction task. Category-based gradient-boosting models achieve moderate performance but do not match the accuracy of the Extreme approach.

These findings underscore the importance of model selection and hyperparameter tuning in predictive analytics for financial time series data. The superior performance of the Extreme Gradient Boosting models suggests that this approach is particularly well-suited to capturing the nonlinear and interactive effects inherent in complaint and sentiment-driven stock price dynamics.

Table 5 compares the out-of-sample predictive performance of four models, highlighting the effects of incorporating complaints, emotions, and advanced machine-learning techniques. Model A, a time-series panel model without complaints and emotions, exhibits the highest prediction error across all metrics (RMSE, MAE and MAPE). This shows the limitations of traditional models that rely solely on financial indicators. Model B, which augments the time-series panel with complaint and emotion variables, achieves modest improvements, indicating that incorporating behavioral factors enhances stock price prediction accuracy.

Applying machine learning through gradient boosting yields a further substantial reduction in prediction errors. Model C (gradient boosting without complaints and emotions) outperforms Models A and B, as evidenced by lower RMSE, MAE and MAPE. Model D, which integrates machine learning and the complete set of complaints and emotion variables, notably achieves the best overall performance. The improvement from Model C to Model D demonstrates the additive value of combining advanced machine learning algorithms with behavioral metrics derived from unstructured text data.

Overall, these findings suggest that the predictive accuracy of stock price models is significantly enhanced by both the inclusion of investor sentiment and complaint information and the adoption of sophisticated machine learning approaches. The results reinforce the importance of behavioral signals and nonlinear modeling for robust financial forecasting.

This study advances both scholarship and practice in financial market analytics by showing how consumer complaint discourse provides an additional perspective on market-relevant sentiment. Prior work on investor sentiment has predominantly relied on investor-centric proxies such as trading volume, flows and survey indices (e.g. Brown and Cliff, 2004; Fisher and Statman, 2000). In contrast, our approach captures market-level patterns in consumer narratives, such as topic concentration and emotional tone, thereby broadening sentiment analysis to include signals drawn from consumer experiences and grievances (Gambetta et al., 2015). This customer-centric perspective complements established sentiment measures by capturing a different segment of public discourse that may become salient to investors and other market participants.

A core methodological contribution lies in integrating NLP-derived indicators with dynamic panel time-series modeling. By applying LDA to extract monthly topic concentration (MTP) and VADER to quantify positive and negative tone, the study converts unstructured complaint narratives into measurable constructs and links them to firm-month stock price outcomes within a unified framework (Blei et al., 2003; Hutto and Gilbert, 2014; Hamilton, 2020). Importantly, because the complaint indicators are aggregated at the month level and merged to firm-month observations by time index, the results should be interpreted as temporal associations between market-level complaint discourse and firm-level stock price dynamics rather than complaint-level causal effects.

The findings extend investor sentiment theory by showing that consumer-generated narratives contain information that is systematically associated with stock prices in the financial services sector (Baker et al., 2012). The positive association between complaint concentration (MTPH) and price suggests that periods in which complaints cluster around a dominant theme coincide with stronger valuation outcomes. One interpretation consistent with limited attention theory is that topic clustering increases salience. When grievances converge on a clearly defined issue, the resulting discourse may attract attention and trigger faster information processing among market participants. This can potentially reflect a “signal” of heightened relevance rather than uniformly negative valuation pressure (Barber and Odean, 2008; Hirshleifer and Teoh, 2003). In this sense, complaint concentration may operate less like a pure “negativity” measure and more like an attention or issue-salience indicator that markets can incorporate in nonobvious ways.

A key clarification is that negative sentiment does not exhibit a robust direct association with stock price in Model B, whereas positive sentiment shows a positive association with price. Accordingly, the theoretical contribution is not that negative emotional expressions directly predict price declines, but rather that negative tone participates in the system through dynamic interdependencies with topic concentration and positive tone. In the structural decomposition, negative sentiment two periods prior is linked to price through indirect pathways. These patterns are consistent with prior work suggesting that affect can shape the informational environment and attention dynamics in ways that influence valuation, though the direction and magnitude of effects may depend on how discourse evolves and is interpreted by the market (Tetlock, 2007; Luo and Homburg, 2008).

The lag structure in the dynamic models indicates that complaint discourse exhibits persistence and feedback. Past concentration predicts subsequent concentration and negative tone, while positive and negative tones show cross-lagged relationships. These results support a “time-ordered association” interpretation. In other words, the results support the hypothesis that shifts in the tone and concentration of complaint discourse are followed by changes in subsequent discourse states and pricing outcomes. However, because the study does not track the resolution trajectory of individual complaints or link each complaint to firm-specific recovery actions, the evidence does not establish a complaint-level causal pathway from “initial negativity” to “post-complaint recovery.” Instead, the findings are best framed as consistent with broader mechanisms proposed in the literature (e.g. acknowledgment, communication or issue containment shaping public discourse) but not as direct tests of those mechanisms (Tetlock, 2007; Luo and Homburg, 2008).

From a market-monitoring perspective, the results suggest that complaint discourse provides incremental information that complements conventional sentiment proxies. Monitoring both topic concentration (as an issue-salience measure) and tone (as an affective measure) may improve early detection of emerging issues and shifts in market mood. This may be useful, particularly in sectors where consumer experiences and trust are central. Practically, investors and risk managers can treat complaint-based indicators as part of an “early-warning dashboard” by using them to flag months with unusually high discourse concentration or sharp polarity shifts and to adjust watchlists, risk limits or hedging practices accordingly (Tetlock, 2007; Gruca and Rego, 2005).

For financial firms, the evidence supports the value of continuously tracking complaint narratives. The narratives should not be limited to volume alone but also to the structure of discourse and emotional tone. Elevated concentration may indicate that complaints are converging on a common failure mode (e.g. credit reporting disputes or account access issues), which can then facilitate faster root-cause triage. Meanwhile, shifts in tone may signal changes in consumer perceptions that merit proactive communication, service recovery, or process redesign. Importantly, because our design captures month-level discourse patterns rather than firm-specific recovery, the managerial implication is not that any single complaint “causes” later valuation changes but that systematic patterns in complaints can help firms prioritize risk mitigation and reputation management promptly (Grönroos, 1984; Gupta et al., 2004; Sirignano and Cont, 2019).

The study also shows that combining dynamic panel structures with ensemble learning can improve predictive performance in complex, noisy financial settings. Gradient boosting appears particularly effective at capturing nonlinearities and interactions among complaint-derived features and fundamentals (Chen and Guestrin, 2016; Zięba et al., 2016). At the same time, the time-series panel component provides a structured interpretation of lagged relationships. This helps mitigate the “black box” critique often raised against purely machine-learning forecasting. For practitioners, a hybrid design can balance predictive gains with clearer economic narratives about persistence, lagged associations and interdependencies.

Overall, the study demonstrates that market-level consumer complaint discourse, measured through topic concentration and sentiment polarity, can be incorporated into dynamic predictive models to improve forecast accuracy while also deepening interpretive insight. The strongest conclusion supported by the evidence is that these complaint-based signals are informative and temporally structured correlates of stock price dynamics. The results encourage further work to refine measurement (e.g. complaint topic taxonomy and alternative sentiment models) and to link complaint discourse more directly to firm-specific response processes where data permit.

Several limitations should be acknowledged when interpreting the findings. First, the complaint-based indicators are constructed from CFPB narratives aggregated at the monthly market (sector) level and then merged into firm-month observations, using the month as the standard time index. As a result, the empirical design supports inferences about time-ordered associations between market-level complaint discourse and firm-level stock prices. However, it does not permit complaint-level causal claims or direct tracking of firm-specific recovery actions following individual complaints. Second, CFPB narratives represent a subset of consumer experiences, and thus the complaint corpus may overrepresent negative experiences and may not capture the full distribution of consumer interactions. Relatedly, because the CFPB database aggregates and anonymizes reports, the analysis cannot reliably connect each complaint to a specific firm's internal response processes, service recovery actions or resolution outcomes. Third, the study operationalizes complaint discourse through LDA-based topic concentration (MTP) and lexicon-based sentiment scoring (VADER). While these methods offer transparency and scalability, they may not fully capture sarcasm, domain-specific language or nuanced contextual meanings in financial complaint narratives. Alternative sentiment models or domain-adapted classifiers could yield different estimates of emotional tone. Fourth, the time-series panel design relies on a fixed monthly aggregation and a specified lag structure. Although the models include extensive controls and lagged dependencies, unobserved shocks (e.g. macroeconomic news, regulatory events or firm-specific disclosures) may co-move with complaint discourse and stock prices, and this can potentially contribute to residual confounding. Finally, the study focuses on publicly traded financial firms, which may limit generalizability to private firms or nonfinancial sectors where complaint dynamics and investor attention operate differently.