Sentiment Analysis on Brazilian Banks on Twitter/x: Comparison between Traditional and Digital Institutions

Fernanda Barberato Leal; Lucas Gabriel da Silva Félix

DOI: 10.22167/2675-6528-202602894

Article derived from a Final Course Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

A study analyzed public perception of Brazilian financial institutions on the Twitter/X platform, highlighting the importance of sentiment monitoring on social networks for understanding reputation and customer experience in the banking sector. The objective was to compare user perception of the image and reputation of traditional and digital banks, based on the sentiment patterns identified in the analyzed manifestations, seeking to identify structural differences between these groups. The methodology was based on the analysis of 1,096 tweets collected between November 2022 and June 2023. Two complementary sentiment analysis approaches were used, the sum and the average of labels, to capture the majority sentiment and nuances of perception. Additionally, the Market Profile Model, with indicators of emotional reputation, reputational risk, neutrality, and polarization, and the Banking Clustering Model, which allowed grouping institutions according to perception patterns, were developed. The results indicated a predominance of neutral and negative sentiments, a higher volume of interactions in digital banks, and structural differences in the emotional intensity of perceptions, with greater stability in digital banks and greater polarization in traditional ones. It was concluded that the combination of analytical and statistical techniques contributed to an in-depth understanding of institutional image in the digital environment, demonstrating the importance of data-driven reputation management strategies.

Keywords: Digital banks; Traditional banks; Data modeling; Opinion mining; Social Networks.

1. Introduction

The growing digitalization of financial services and the popularization of social media have transformed how consumers interact with banking institutions (Kotler et al., 2017). Platforms like Twitter/X have become relevant spaces for expressing opinions, complaints, and praise about products and services, allowing for real-time monitoring of public perception of brands and companies.

In the Brazilian banking context, digital transformation has modified the competitive dynamics of the sector and expanded the possibilities for relationships between financial institutions and clients. The expansion of fintechs and digital banks has introduced new service delivery models, driving traditional banks to incorporate digital solutions into their operations (Pinto et al., 2023). In this environment of growing competition and transformation, user perception becomes a relevant element for understanding the image and reputation of financial institutions, as the experiences and opinions publicly expressed on social media can provide evidence of how these organizations are perceived in the digital environment (Pinto et al., 2023).

Sentiment analysis, also known as opinion mining, is a field of text mining that seeks to automatically identify the polarity (positive, negative, or neutral) of opinions expressed in natural language (Liu, 2012). This approach combines natural language processing (NLP) and machine learning techniques to extract subjective information from large volumes of text, being widely used in contexts of social media monitoring and corporate reputation studies (Pang and Lee, 2008).

Several studies have applied sentiment analysis to data from social platforms to understand consumer perceptions across different sectors. In the specific case of the financial sector, recent research indicates that monitoring public sentiment can assist institutions in strategic decision-making and early identification of image crises (Pak, 2010). In this regard, the present study proposes to investigate and compare the sentiment expressions of Twitter users about Brazilian banks, using the Brazilian Banks Tweets (B2T) database (Kakimoto et al., 2024).

Beyond sentiment analysis, the study incorporated clustering techniques as a way to deepen data interpretation and identify structural patterns of public perception among the analyzed institutions. Clustering consists of an unsupervised model technique aimed at forming groups of elements that are similar to each other and distinct from other groups, allowing for the segmentation of observations based on common characteristics. Among existing methods, the K-Means algorithm stands out as one of the most widespread approaches for numerical data, being widely used in identifying homogeneous patterns and market segmentation (MacQueen, 1967; Han et al., 2012). In comparative analyses, K-Means presents advantages related to computational simplicity, operational efficiency, and high interpretability of the formed groups, allowing for the identification of distinct behavior and perception profiles (Hair Jr. et al., 2014).

Thus, the combination of sentiment analysis and clustering sought to offer a more comprehensive quantitative approach to the study of the institutional banking image. This allows not only for the measurement of emotional polarities but also for the identification of strategic segments and comparative patterns between traditional and digital banks. The research addresses the problem of understanding how the structural changes in the financial system, especially the digitalization of services, influence the public perception of banking brands.

The importance of this study lies in the possibility of understanding the dynamics of public perception in the digital environment and in the contribution to the field of sentiment analysis applied to the Brazilian context and to corporate reputation management in the financial sector. Given this context, this research aims to analyze, comparatively and quantitatively, the sentiments expressed by Twitter/X users in relation to Brazilian banking institutions, considering the distinction between traditional banks and digital banks, seeking, from the identified sentiment patterns, to understand differences in user perception and their possible implications for the image and reputation of institutions in the digital environment.

2. Material and Methods

This research was characterized as a quantitative and descriptive study, conducted in the form of a multiple case study. Different traditional and digital banking institutions operating in Brazil were analyzed and compared. The main objective of the descriptive approach was to describe the characteristics of public perception of banks on the Twitter/X platform, as defined by Gil (2019).

Additionally, the research was configured as documentary, since it used secondary data already publicly available. The quantitative approach was justified by the application of statistical and computational techniques for the measurement, comparison, and interpretation of user sentiments, aligning with the objective of comparatively analyzing the perception of traditional and digital banks.

The database used was the Brazilian Banks Tweets (B2T), developed by Kakimoto et al. (2024) and made publicly available in a GitHub repository. This dataset consisted of 1,096 tweets collected from the Twitter/X platform, covering the period from November 2022 to June 2023. The data included the tweet ID, publication date and time, language, name of the bank mentioned, and sentiment labels (positive, negative, and neutral) assigned by volunteers.

The banks included in the analysis were Itaú, Bradesco, Santander, Banco do Brasil, BTG Pactual, NuBank, Banco Inter, Banco Pan, Sofisa Direto, and Banco BRB. The user’s location information, present in the database, was not used due to the large number of null records and its imprecision, focusing on the other available variables for the analysis.

The collection and analysis procedures began with importing the B2T database, in CSV format, into the Google Colab environment. After loading the dataset, data reading and standardization were performed, including formatting the date and time field and creating a field named ‘anomes’ to facilitate temporal analysis.

Given that the B2T dataset did not contain the original text of the tweets, in compliance with Twitter/X data sharing guidelines, the analysis was conducted based on the sentiment labels previously assigned to the records. As described by Kakimoto et al. (2024), the tweets were subjected to a manual labeling process carried out by five volunteers, who independently classified the comments into three categories: positive (+1), neutral (0), or negative (-1).

Each tweet was evaluated by at least three volunteers, without access to the ratings of other reviewers. The definition of sentiment classes followed criteria established by the database authors, where the positive classification (+1) indicated satisfaction or praise, negative (-1) expressed discontent or dissatisfaction, and neutral (0) was used for expressions without clear polarity, such as factual or incomprehensible texts.

From the multiple labels assigned to each tweet, two complementary measures were calculated for the final sentiment classification: the sum and the average of the labels. The sum of the labels represented the polarity balance of the reviews, indicating the predominant direction of the classifications. The average of the labels, in turn, was used as a normalized measure of polarity, allowing for consideration of the intensity and direction of the reviews regardless of the number of reviewers.

For the final classification of tweets, the rule was adopted that positive values resulting from the sum or average were classified as positive (+1), negative values as negative (-1), and values equal to zero as neutral (0). These approaches allowed for the production of quantitative and comparable classifications of predominant sentiments, which were subsequently used in the aggregated analyses by banking institution and period.

The data analysis stage was conducted descriptively and comparatively, with the objective of identifying patterns and trends in the sentiments expressed by users. Initially, a descriptive analysis of the results was performed, involving the calculation of absolute and relative frequencies, as well as the identification of proportions and trends of the sentiments observed in each group, considering dimensions such as banking institution, reference month, and type of institution (traditional or digital).

In the comparative stage, the banks were classified into two groups: traditional and digital. Traditional banks were considered those with a consolidated trajectory in the Brazilian banking system and historically structured operations in physical channels (Bradesco, Itaú, Santander, Banco do Brasil, and Banco BRB). Digital banks were classified as those whose analyzed operations showed a greater predominance of digital channels and services (NuBank, Banco Inter, Banco Pan, Sofisa Direto, and BTG Pactual).

Additionally, the Market Profile Model (MPM) was developed, with the objective of synthesizing user perception through derived indicators. This model was composed of the indicators of Emotional Reputation (RE), Neutrality (N), Emotional Polarization (PE), and Reputational Risk (RR). These indicators were calculated from the proportions of positive, negative, and neutral sentiments observed for each institution.

Emotional Reputation (ER) was defined as the difference between the proportion of positive sentiments and the proportion of negative sentiments. Neutrality (N) corresponded to the proportion of tweets classified as neutral. Emotional Polarization (EP) was calculated by summing the proportions of positive and negative sentiments. Reputational Risk (RR) was calculated as a composite measure by the interaction between the proportion of negative sentiments and emotional polarization.

Complementarily, the Banking Clustering Model (BCM) was applied, with the objective of identifying structural patterns in public perception. For this, the proportions of positive and negative sentiments were used as input variables in an unsupervised model algorithm, K-Means, allowing the segmentation of banks into clusters with similar characteristics. The K-Means algorithm was adopted for its suitability for continuous numerical data and for exploratory segmentation analyses (MacQueen, 1967; Han et al., 2012; Hair Jr. et al., 2014).

After defining the K-Means algorithm, different values of K, ranging from two to six clusters, were tested to identify a configuration that presented adequate structuring of the groups. For this definition, the Elbow Method and the Silhouette Index were used jointly. The solution with three clusters was adopted as the final configuration of the clustering model, as it presented a balance between the reduction of internal variability and the quality of grouping, in addition to greater interpretability of the groups in the context of the analysis.

3. Results and Discussion

The analysis of data from the Brazilian Banks Tweets (B2T) database, which comprised 1,096 tweets collected between November 2022 and June 2023, revealed significant patterns in public perception of Brazilian financial institutions. The study employed two complementary approaches for sentiment analysis: the sum of labels and the average of labels, allowing for a multifaceted understanding of user expressions. These approaches were crucial for identifying both the majority sentiment and the intensity nuances in perception, providing a robust basis for comparisons between traditional and digital banks, as well as for modeling market profiles and banking clustering.

Initially, the general distribution of sentiments, using the sum of labels, indicated a predominance of neutral classifications, representing 46.4% of the total tweets. Negative sentiments corresponded to 42.5%, while positive ones were the least frequent, with 11.0%. This approach emphasizes the net balance of evaluations, where the compensation between positive and negative votes can result in a neutral classification, even if the tweet contains elements of polarity. Such a characteristic suggests that a considerable portion of interactions does not express a clear or univocal polarity, but rather a combination of perceptions that mutually cancel each other out.

In contrast, the label averaging approach presented a different distribution, with 55.4% of tweets classified as negative, 25.4% as neutral, and 19.3% as positive. Averaging, by expressing polarity on a continuous scale, allowed for capturing the relative intensity of the trend of evaluations, even in cases of divergent opinions. This difference between the approaches highlights the importance of using multiple metrics to characterize user perception, preventing interpretation from being restricted to a single aggregation method and revealing that the intensity of negative sentiment is more pronounced when considering the average of the evaluations.

When analyzing the distribution of sentiments by bank type, it was observed that digital banks concentrated the largest volume of tweets in the sample. For both groups, the predominance of sentiments varied according to the approach. In the sum of labels, digital banks showed a higher concentration of neutral tweets, while in traditional banks, negative ones prevailed. However, in the average of labels approach, negative sentiment was predominant in both groups. In all analyses, the proportion of positive sentiments remained the lowest, indicating that public perception tends to be more critical or neutral, regardless of the banking operational model.

The individual analysis by financial institution revealed that NuBank stood out for the highest volume of tweets, representing 38% of the total sample. For this bank, there was a higher concentration of negative and neutral sentiments, depending on the approach used. Traditional banks such as Itaú, Bradesco, and Santander also presented a predominance of negative and neutral tweets, reflecting more balanced user opinions. Institutions with lower representation in the sample, such as Sofisa Direto and BTG Pactual, showed reduced volumes of mentions, which limits the interpretation of specific sentiment patterns for these institutions, but still highlight that the type of bank influences public perception.

The temporal analysis of sentiments, aggregating data by month and year, showed a growth in tweet volume starting in November 2022, reaching a peak in January 2023. During this period, an increasing concentration of negative sentiments was observed. This trend may be associated with specific campaigns or year-end/start events that impacted users’ perception of financial institutions. The evolution of sentiments over time suggests that banking reputation in the digital environment is dynamic and sensitive to contextual factors, requiring continuous monitoring to identify changes and respond proactively.

The analytical modeling was developed to synthesize user perception through derived indicators, such as Emotional Reputation, Neutrality, Emotional Polarization, and Reputational Risk. These indicators were calculated from the proportions of positive, negative, and neutral sentiments for each institution, allowing for a multidimensional comparison between banks. Emotional Reputation (ER) was defined as the difference between the proportion of positive and negative sentiments, ranging from -1 to 1, where positive values indicate a predominance of favorable sentiments and negative values indicate unfavorable ones. Neutrality (N) corresponded to the proportion of neutral tweets, indicating the absence of emotional polarity. Emotional Polarization (EP) was the sum of the proportions of positive and negative sentiments, representing the degree of emotional intensity. Finally, Reputational Risk (RR) was calculated as the interaction between the proportion of negative sentiments and emotional polarization, indicating potential exposure to image crises.

The Banking Clustering Model (BCM) was applied using the K-Means algorithm, with the proportions of positive and negative sentiments as input variables. Three distinct clusters were identified, reflecting different public perception profiles. The first cluster, composed of Bradesco, BTG Pactual, Banco Pan, and BRB, was characterized by the predominance of negative evaluations and low compensation for positive perceptions, indicating greater reputational vulnerability. These banks, from a strategic point of view, are more exposed to risks related to institutional image, suggesting the need for interventions to improve public perception and reduce recurring dissatisfactions.

The second cluster included Itaú, Banco do Brasil, and Nubank, presenting a more balanced profile, with a greater relative weight of positive perceptions, which contributes to a more favorable institutional image. The presence of Nubank, a digital bank, alongside traditional banks, suggests that, despite structural differences in business models, some traditional institutions manage to achieve perception levels comparable to digital ones. This cluster represents banks with a more robust competitive positioning in terms of emotional reputation, indicating greater user acceptance.

The third cluster was formed exclusively by Banco Sofisa Direto, characterized by low emotional intensity in interactions, with users who do not express clear opinions about the institution. This behavior may be associated with lower brand visibility, a reduced volume of interactions, or a more functional and less experiential relationship with the bank. Although low negativity may seem positive, the absence of positive feelings also suggests limitations in building a strong and differentiated institutional image, indicating a strategic challenge in terms of brand positioning and recognition. Inter and Santander banks positioned themselves in an intermediate zone, suggesting a public perception in transition, with a coexistence of negative and positive evaluations, indicating instability.

The Market Profile Model (MPM) showed clear differences between traditional and digital banks. Both groups presented negative Emotional Reputation, with traditional banks at -0.37 and digital banks at -0.33, indicating that, although both have a negative balance, traditional banks are relatively more negative. Regarding Reputational Risk, traditional banks scored 0.24, slightly higher than the 0.21 of digital banks, suggesting a slightly higher combination of negativity and emotionally polarized expressions in traditional banks. Emotional Polarization was 0.53 for traditional banks and 0.43 for digital banks, indicating that 53% of expressions in traditional banks and 43% in digital banks showed some emotional polarity.

Complementarily, the Neutrality indicator revealed that 57% of the manifestations about digital banks were neutral, compared to 47% in traditional banks. The greater neutrality in digital banks indicates a higher proportion of manifestations without clear polarity, which should not be interpreted as a necessarily more positive perception, but rather as an absence of polarity. These findings corroborate the literature that points to differences in user perception of digital and traditional banks, especially in dimensions such as utility, convenience, interaction, and security (Shin, Cho, and Lee, 2020), reinforcing the existence of differentiated characteristics in user perception according to the banking model.

In summary, the results demonstrate that the type of bank significantly influences public perception, that the methodological approach impacts the interpretation of sentiments, and that there are structural differences in both the intensity and nature of user interactions with financial institutions. The integration of sentiment analysis, market profile modeling, and clustering techniques proved effective in broadening the understanding of institutional image in the digital environment. Traditional banks showed greater polarization and reputational risk, while digital banks exhibited greater neutrality and lower polarization, characterizing distinct profiles of public perception and evidencing the utility of quantitative sentiment analysis for reputation management.

4. Conclusion

This study comparatively and quantitatively analyzed the sentiments expressed by Twitter/X users regarding Brazilian banking institutions, distinguishing between traditional and digital banks, with the aim of understanding user perception differences and their implications for image and reputation in the digital environment. A predominance of neutral and negative sentiments was observed in the analyzed manifestations, although the exact distribution varied according to the methodological approach employed. Digital banks concentrated a higher volume of interactions, and individual analysis highlighted NuBank with the largest number of records. The application of the Market Profile Model (MPM) revealed that traditional banks exhibited greater emotional polarization and reputational risk, in addition to a more negative emotional reputation. In contrast, digital banks presented greater neutrality and lower polarization and reputational risk, with a less negative emotional reputation. Additionally, the Banking Clustering Model (MCB) identified three distinct profiles of public perception, ranging from institutions with high reputational vulnerability to those with a more balanced profile and others with low emotional intensity. These findings highlight significant structural differences in the intensity and nature of user interactions with financial institutions, providing an in-depth understanding of institutional image in the digital environment and contributing to the formulation of data-driven reputation management strategies.

Despite the methodological robustness, the study has important limitations, such as the absence of the original textual content of the tweets in the database used, which prevented a more in-depth qualitative analysis of the reasons underlying the expressed sentiments. Furthermore, the reduced volume of mentions for some institutions, such as Sofisa Direto and BTG Pactual, restricted the interpretation of specific sentiment patterns. For future studies, it is suggested to investigate the contextual factors and specific events that may influence sentiment trends over time, especially during periods of high polarization. It is also recommended to explore methodologies that allow for the analysis of textual content, if available, to uncover the causes of perceptions and deepen the understanding of user engagement dynamics. The observed instability in the perception of banks like Inter and Santander also points to the need for research that explores the factors leading to this transition of sentiments.

Bibliographic References

Gil, A. C. (2019). Métodos e Técnicas de Pesquisa Social. 7. ed. São Paulo: Atlas.

Hair Jr., J. F. et al. (2014). Multivariate Data Analysis. 7. ed. Harlow: Pearson.

Han, J.; Kamber, M.; Pei, J. (2012). Data Mining: Concepts and Techniques. 3. ed. Burlington: Morgan Kaufmann.

Kakimoto, G. K.; Haddadi, S. J.; Araújo, P. M.; Silva, F. S.; Reis, J. C.; Reis, M. S. (2024). B2T: A Dataset of Tweets in Portuguese Language about Brazilian Banks. Proceedings of the VI Dataset Showcase Workshop (DSW). Disponível em https://github.com/GabrielKakimoto/B2T-A-Brazilian-Tweet-Banks-Dataset-in-Portuguese-Language . Acesso em: 29 out. 2025.

Kotler, P.; Kartajaya, H.; Setiawan, I. (2017). Marketing 4.0: Moving from Traditional to Digital. Coimbra, Portugal: Conjuntura Actual Editora. Trad. Pedro Elói Duarte.

Liu, B. (2012). Sentiment Analysis and Opinion Mining. Morgan & Claypool Publishers.

MacQueen, J. (1967). Some methods for classification and analysis of multivariate observations. In: Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability. Berkeley, p. 281-297.

Pak, A.; Paroubek, P. (2010). Twitter as a Corpus for Sentiment Analysis and Opinion Mining. LREC.

Pang, B.; Lee, L. (2008). Opinion Mining and Sentiment Analysis. Foundations and Trends in Information Retrieval. Vol.2.

Pinto, A. R.; Martens, C. D. P.; Kniess, C. T.; Oliveira Filho, B. G. (2023). Digital entrepreneurship in the Brazilian banking sector: analysis based on the emergence of fintechs. Future Studies Research Journal: Trends and Strategies, Vol. 15 No. 1.

SHIN, Jae Woo; CHO, Ji Yeon; LEE, Bong Gyou. (2020). Customer perceptions of Korean digital and traditional banks. International Journal of Bank Marketing, v. 38, n. 2, p. 529–547. DOI: 10.1108/IJBM-03-2019-0084.

Article originating from the Final Course Work of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform