“Churn” Prediction in “Igaming” Platforms using “Machine Learning” and “Marketing Analytics”

Douglas Caldeira Silva; Nuno Manoel Martins Dias Fouto

DOI: 10.22167/2675-6528-202602832

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

The prediction of “churn” in “iGaming” platforms has proven relevant due to the non-contractual nature of this environment, where churn is inferred from observable behavioral and transactional signals. The work aimed to develop and evaluate predictive models capable of estimating the probability of user “churn” on “iGaming” platforms, based on behavioral and transactional variables, to support marketing analytics decisions focused on anticipating churn. An applied, quantitative, predictive, and retrospective observational study was conducted, using secondary data from a company in the sector. The analytical base was structured in a daily user panel, with variables derived from recency, frequency, intensity of use, and relationship maturity. The operational definition of “churn” considered inactivity windows of 15 and 30 days, adopting the 30-day window as the main scenario. Logistic Regression and Random Forest models were employed and evaluated. The results indicated that both algorithms presented high predictive performance. Logistic Regression stood out as the main alternative by combining high accuracy (0.8980), better performance in AUC-ROC, AUC-PR, and precision, in addition to greater interpretability, while Random Forest achieved slightly higher recall and equivalent F1-score. It was concluded that the use of machine learning techniques is viable to support segmentation, user prioritization, and retention actions in “iGaming”, provided it is accompanied by continuous validation, performance monitoring, and periodic review of the operational definition of “churn”.

Keywords: Non-contractual environment; Customer churn; Random Forest; Logistic regression; Customer retention.

1. Introduction

Digital transformation has boosted organizations’ ability to collect, store, and analyze data on consumer behavior, making decisions increasingly evidence-based. In this scenario, marketing analytics assumes a strategic position, assisting in identifying consumption patterns, audience segmentation, and offer personalization, as well as in developing actions for customer acquisition, retention, and value maximization. Erevelles et al. (2016) highlight that technological advancements have expanded the volume, velocity, and variety of available data, creating new possibilities for understanding consumer behavior and guiding marketing decisions. Complementarily, Kumar and Reinartz (2018) emphasize the contribution of CRM and marketing analytics to segmentation, campaigns, retention, and customer value management. Within this context, the phenomenon of churn, or customer attrition, emerges as one of the main challenges for companies in digital markets, as it is directly associated with the interruption or weakening of the customer relationship with the organization (Jahromi et al., 2014).

In the iGaming sector, this discussion acquires particular relevance due to the segment’s characteristics, which encompass digital platforms for sports betting, online casino, and poker, operating in a highly competitive, dynamic environment dependent on users’ recurring engagement. In these markets, customer retention is crucial, as competition is intense, migration between platforms can occur easily, and usage recurrence directly impacts business performance. The ability to identify early signs of churn becomes, therefore, fundamental for organizations seeking to mitigate revenue losses, increase user retention, and optimize their relationship strategies. Merchie and Ernst (2022) reinforce that churn prediction in online gambling is compatible with predictive approaches based on users’ activity history.

Literature indicates that customer churn generally does not occur randomly, often being preceded by observable changes in customer behavior, such as variations in the recency, frequency, and monetary value of interactions. Kaya et al. (2018) demonstrate that dynamic behavioral patterns, extracted from transactional records, can predict churn decisions more effectively than demographic attributes. Mena et al. (2024) complement this by stating that time-varying measures of recency, frequency, and monetary value constitute a relevant basis for churn prediction. This perspective is even more critical in non-contractual business models, where customer loss does not manifest as an explicit cancellation but through observable signs of reduced or interrupted activity. Fader and Hardie (2007) explain that in non-contractual contexts, the moment a customer becomes inactive is not directly observable by the company, which brings the phenomenon of churn closer to the reality of iGaming platforms, where evasion is inferred from user transactional behavior.

In Brazil, the importance of the topic was intensified by the recent regulation of the fixed-odds betting and online gaming market. Law No. 14.790, of December 29, 2023, established the regulatory framework for the exploitation of fixed-odds betting and included online gaming, conditioning the operation on prior authorization from public authorities (BRAZIL, 2023). Additionally, the Ministry of Finance registered the start of full operation of the regulated market in January 2025 (BRAZIL, 2024), which reinforces the sector’s institutionalization and the need for more structured management practices. In a regulated, competitive, and data-intensive environment, understanding user behavior and anticipating evasion movements become even more relevant for marketing and management decisions.

The choice of theme also aligns with the growing use of data in organizational decision-making. With the greater availability of behavioral and transactional records, it is possible to employ analytical methods to identify patterns and classify user profiles, supporting evidence-based decisions. Mitchell (1997) explains that machine learning is based on the construction of systems that improve their performance from experience, offering theoretical support for its application in problems with large volumes of data and pattern recognition. In the context of iGaming, this approach is particularly suitable, as churn does not always manifest as an explicit cancellation; in many cases, the user simply reduces their activity or stops interacting with the platform, requiring that abandonment be identified by signals observed in the user’s behavior itself.

Given this scenario, the following research question emerges: to what extent do behavioral and transactional variables allow predicting churns in iGaming platforms, in a non-contractual environment? From an academic perspective, the study integrates the fields of marketing analytics, customer retention, and machine learning, aligning with Data Science. From a practical point of view, the research can contribute to the early identification of users at higher risk of churn, favoring the prioritization of data-driven retention, segmentation, and relationship actions. In this context, the general objective of this work consists of developing and evaluating predictive models capable of estimating the probability of churn of users on iGaming platforms, based on behavioral and transactional variables, in order to support marketing analytics decisions aimed at anticipating churn.

2. Material and Methods

The research was characterized as applied, with a quantitative approach, predictive purpose, and retrospective observational design. The predictive objective consisted of estimating the probability of “churn” in iGaming platforms, using structured data on user behavior and transactions, analyzed by statistical and machine learning techniques (Gil, 2002; Prodanov and Freitas, 2013). For this, secondary data from a large company in the iGaming sector were used, whose identity was preserved due to confidentiality. The unit of analysis was the user-date instance, structured in a daily panel per user. The initial data extraction was performed through SQL queries on the `kpis_usuarios_daily` base, gathering variables such as deposits, withdrawals, “cashback”, casino “turnover”, casino bonus, and gross gaming revenue (GGR), related to the start of the relationship, transactional volume, and platform activity intensity.

For the construction of the analytical base, a cohort of users who made their first deposit (FTD) in January 2025 was defined, followed throughout the year. The complete daily panel per user was constructed to explicitly represent days with and without activity, essential for the labeling of “churn” in a non-contractual environment. The base underwent consistency and preparation procedures, including data standardization, duplicate checking, and inconsistent date verification. Observations with FTD after the reference date and instances without a sufficient future window for labeling the target variable were excluded to avoid temporal leakage.

The target variable, “churn”, was inferred based on the observable interruption of user transactional activity, considering financial movement (deposit or withdrawal) as relevant activity. Windows of 15 and 30 consecutive days of inactivity were evaluated, with the 30-day window being adopted as the main scenario due to being more conservative and reducing the label’s sensitivity to temporary pauses.

The variables included in the modeling were selected for their empirical availability and theoretical adherence to the “churn” problem, prioritizing attributes of recency, frequency, monetary intensity, relationship maturity, and temporal changes in movement patterns (Kaya et al., 2018; Mena et al., 2024). Among the main ones, `valor_ftd_usuario`, `tempo_desde_ftd`, `ticket_medio_deposito_30d`, `desaceleracao_atividade`, `dias_ativos_15d`, `qtd_deposito_15d`, and `proporcao_dias_ativos_30d` stood out, which synthesized the user’s recent activity.

For “churn” prediction, supervised classification algorithms were used. Logistic Regression was adopted as the main and baseline model, due to its simplicity and interpretability (James et al., 2021). The Random Forest algorithm was employed as a comparative model, for its ability to capture non-linear relationships (Breiman, 2001).

The validation of the models was performed through temporal holdout, a strategy suitable for datasets with a chronological structure (Roberts et al., 2017). The temporal cutoff for analytical separation was September 25, 2025, ensuring that the predictors were calculated with information available up to the reference date.

In the conservative specification, Logistic Regression was trained with 739,349 observations and tested on 197,232. The Random Forest was trained with 738,995 observations and tested on 196,979. Predictive performance was analyzed by “accuracy”, AUC-PR, AUC-ROC, precision, “recall”, “F1-score”, and confusion matrix (Fawcett, 2006; Powers, 2011; Saito and Rehmsmeier, 2015).

The computational procedures were performed in a Databricks environment. Initial data extraction was done through SQL queries, while base preparation, variable creation, and target variable definition were conducted in PySpark. Modeling utilized machine learning libraries. The study was developed with secondary data from an iGaming company, with its identity preserved due to confidentiality, focusing on transactional and temporal behavioral patterns, without the objective of individual user identification, respecting informational secrecy.

3. Results and Discussion

This section presents the results obtained in the empirical stage of the study, with an emphasis on the definition of the target variable, the characteristics of the analytical base, and the comparison between the Logistic Regression and Random Forest algorithms in the main scenario of 30-day “churn”. The analysis was conducted based on the results observed in the execution of the models in the Databricks environment, seeking to interpret the findings not only in statistical terms but also in terms of their analytical and managerial utility for marketing analytics in iGaming. The discussion articulates predictive performance, the error profile of the classifiers, and the operational implications for retention actions in a non-contractual environment, offering a comprehensive understanding of the application of machine learning models in this context.

The target variable, representing “churn”, was constructed based on the absence of future user activity in a non-contractual environment. Initially, two operational “churn” scenarios were evaluated: 15 and 30 consecutive days without a deposit or withdrawal. In the 15-day scenario, the positive “churn” class represented 79.58% of the instances, while the negative class corresponded to 20.42%. When extending the inactivity window to 30 days, the positive class came to represent 73.90% of the observations and the negative class 26.10%. This behavior indicates that adopting a longer window reduces the label’s sensitivity to short-term temporary pauses.

The 30-day definition proved to be more conservative for analytical purposes and was adopted as the main scenario for comparing algorithms, although the choice remains operational and dependent on the business context. This longer window is more aligned with the typical intermittent behavior of digital iGaming platforms. In non-contractual environments, the mere absence of activity for a few days does not necessarily represent effective abandonment, as it may reflect natural pauses in usage patterns. Thus, the 30-day horizon tends to produce a more stable label for modeling and is more compatible with retention decisions that require distinguishing between temporary fluctuation and probable churn (Fader et al., 2010; Merchie and Ernst, 2022).

From a descriptive point of view, the analytical base showed a strong concentration of values equal to zero in daily transactional variables, such as deposits, withdrawals, “turnover”, and GGR. This pattern is consistent with the user-date structure used in the study and with the discontinuous nature of user activity on iGaming platforms. Analytically, this means that an important part of the base consists of days with no activity, which does not constitute a construction error, but a substantive characteristic of the observed phenomenon. In contrast, after refining the variables, the `valor_ftd_usuario` attribute began to more consistently represent a structural characteristic of the user throughout their journey, expanding the informational utility of the “first-time deposit” for modeling.

This descriptive behavior already suggests, even before the predictive stage, that the identification of “churn” tends to depend less on a single isolated monetary event and more on the combination of intensity, frequency, and recency of use. Descriptive statistics indicate that the base combines two relevant properties for the “churn” problem: high asymmetry in monetary variables and a strong presence of daily inactivity. In terms of marketing analytics, this pattern is particularly important because it indicates that the loss of connection with the platform does not stem solely from an abrupt drop in large transactional values, but also from gradual changes in frequency and intensity of use.

In the process of building the base, two aspects were particularly relevant for the temporal coherence of the problem. Firstly, the creation of the complete daily panel per user allowed for the explicit representation of days without activity, a necessary condition for the correct labeling of “churn” in a non-contractual environment. Secondly, the target variable was always constructed from a future observation window, in order to avoid the use of information subsequent to the prediction moment. In the main scenario of “churn” in 30 days, the analytical separation was performed with a temporal cutoff on September 25, 2025.

This split resulted in 739,349 observations in training and 197,232 in testing for conservative Logistic Regression, and 738,995 observations in training and 196,979 in testing for conservative Random Forest. Methodologically, this care is central, as the utility of a churn model depends on its ability to operate with information available up to the reference date, and not on retrospectively reproducing a behavior already consummated. The interpretation of the results must consider that, in problems of this nature, the robustness of the evidence also depends on the temporal distance between observation and event and on additional out-of-sample validations.

The comparison between the algorithms was performed in the main scenario of 30-day “churn”. Logistic Regression was treated as the main conservative model due to its combination of high predictive performance with greater interpretability, while Random Forest was used as a comparative model, allowing verification of whether a non-linear tree-based approach would yield substantial performance gains. This choice also allows evaluating whether the problem structure demands greater algorithmic complexity or if a relevant part of the signal can be captured by a simpler and auditable classifier. The conservative nature of the comparison is especially relevant because, in the main specification, variables excessively close to the operational definition of “churn” were excluded.

The conservative Logistic Regression presented an accuracy of 0.8980, an AUC-ROC of 0.9212, and an AUC-PR of 0.9743. Additionally, it achieved a precision of 0.9232, a recall of 0.9540, and an F1-score of 0.9384. These values indicate robust predictive performance, with a high capacity to discriminate between churn and non-churn classes at different decision thresholds. The combination of high metrics reinforces the model’s viability for anticipating churn on iGaming platforms, offering significant support for marketing analytics decisions.

The conservative Random Forest, in turn, obtained an “accuracy” of 0.8976, an AUC-PR of 0.9641, an AUC-ROC of 0.9044, a “precision” of 0.9127, a “recall” of 0.9664, and an “F1-score” of 0.9388. In strictly numerical terms, the differences between the models were small. This result suggests that a substantial part of the predictive signal can be captured even by a simpler decision boundary. However, the proximity between performances also recommends caution: in problems with very informative recent variables, high results can reflect both the anticipation capacity and the strong proximity between the observed behavior and the event one seeks to predict.

The reading of the metrics needs to be done with caution, especially due to the asymmetric distribution of classes. As the positive class of “churn” represents 73.90% of the base in the 30-day scenario, “accuracy”, in isolation, has limited utility as a decision criterion. As a minimum reference, a naive classifier that always predicted “churn” would already achieve an accuracy of approximately 0.7390, without offering comparable discriminatory capacity. In this context, conservative Logistic Regression raised the “accuracy” to 0.8980, which represents a gain of about 15.9 percentage points over this trivial baseline.

For this reason, the most relevant metrics for the interpretation of the results are AUC-PR, AUC-ROC, “precision”, “recall”, “F1-score”, and, complementarily, the confusion matrix. From this perspective, Logistic Regression showed better performance in AUC-ROC, AUC-PR, and “precision”, while Random Forest achieved slightly higher “recall” and practically equivalent “F1-score”. This pattern indicates that Random Forest was slightly more aggressive in recovering actual “churns”, whereas Logistic Regression better preserved the separation between classes across thresholds and the relative control of false positives.

The analysis of the confusion matrix reinforces this profile difference. In the consolidated execution, the conservative Logistic Regression produced 23,931 true negatives, 12,691 false positives, 7,350 false negatives, and 152,553 true positives. The conservative Random Forest, in turn, produced 22,207 true negatives, 14,781 false positives, 5,376 false negatives, and 154,479 true positives. In substantive terms, this means that the Random Forest recovered more real “churns”, but at the cost of incorrectly triggering a larger number of users.

As these matrices depend on the classification threshold adopted during the experiments, they should be read as an operational portrait of the “trade-off” between coverage and selectivity in the tested scenario. In business terms, this difference is relevant because the choice of the model depends not only on the aggregated statistical performance but also on the operational logic of retention actions. In a context where the cost of contacting or rewarding users is high, a model with higher “precision” may be more suitable. On the other hand, in strategies where the priority is to maximize the coverage of users at risk of evasion, a model with higher “recall” may be preferable.

This interpretation brings the results closer to the field of marketing analytics. In practical applications, a “churn” model is not used solely for abstract classification, but to guide decisions such as campaign prioritization, CRM audience definition, risk segmentation, and promotional budget allocation. In this sense, the difference between algorithms can be seen as a difference between activation strategies: one more selective, with less dispersion of efforts, and another more comprehensive, with greater coverage, but with a higher risk of unnecessary contacts.

From a managerial perspective, this means that the predictive score can be incorporated into distinct relationship strategies, combining attrition risk with user value, promotional sensitivity, and expected intervention cost. This interpretation aligns with studies that position “churn” prediction as a relevant tool for retention and decision-making in CRM within the online gambling sector (Coussement and De Bock, 2013). The analysis of variable importances in the Random Forest model highlighted the predominance of attributes associated with recent activity and transactional frequency.

In the case of Random Forest, among the six most relevant attributes, `dias_ativos_15d`, `qtd_deposito_15d`, `qtd_deposito_30d`, `deposito_15d`, `proporcao_dias_ativos_30d`, and `deposito_30d` stood out. These findings indicate that the distinction between users with a higher and lower probability of “churn” is strongly associated with the intensity and recent recurrence of activity. This result is consistent with the literature on “churn” in non-contractual contexts, which tends to highlight the centrality of indicators related to recency, frequency, and behavioral intensity (Fader et al., 2010).

In the specific case of iGaming, the relevance of these variables suggests that the weakening of the link with the platform becomes observable primarily through the reduction of transactional activity in the windows closest to the reference date, in line with the formulation of “churn” as a classification problem in online gambling presented by Merchie and Ernst (2022). At the same time, the predominance of very recent attributes imposes an important caveat: part of the performance may stem from the fact that such variables function as markers very close to the “churn” event itself.

This strengthens short-term prediction, but does not eliminate the need for additional tests to assess generalization over longer horizons. In managerial terms, this means that predictive models can be particularly useful for identifying early signs of disengagement, allowing users to be triggered before churn consolidates. However, the findings suggest greater utility for short-term interventions than for long-horizon strategic decisions, as the strength of recent variables tends to reduce the distance between the detection of behavioral weakening and the materialization of “churn”.

Thus, although the results are favorable to the use of machine learning for predicting “churn”, they should be understood within the specific scope of this study, without excessive generalizations and without assuming equivalent performance in other time windows or in databases structured differently. Regarding the interpretability of Logistic Regression, the model’s coefficients represent the change in the log of the odds of “churn” associated with each variable, with all other variables held constant. Positive coefficients indicate an association with an increased chance of “churn”, while negative coefficients indicate an association with a reduction in these chances.

Exponentiation of the coefficient provides the odds ratio: values above 1 indicate an increase in relative chances and values below 1 indicate a reduction. However, the consolidated version of the modeling used in this work did not preserve the table with the individual coefficients of the variables. For this reason, specific values that are not recorded in the original results are not presented. The substantive interpretation of the Logistic Regression therefore remains supported by the performance metrics, the observed error profile, and the behavioral logic of the variables used in the modeling.

The choice of Logistic Regression as the main model is defensible in the analyzed context. Although Random Forest presented a slightly superior “recall” and practically equivalent “F1-score”, Logistic Regression achieved better results in AUC-ROC, AUC-PR, and “precision”, in addition to offering greater interpretability. This pattern suggests that the introduction of greater algorithmic complexity did not produce proportional gains in the metrics most useful for a selective managerial application. In practical terms, the difference between the models is equivalent to favoring greater coverage of real “churns”, in the case of Random Forest, or a lower volume of undue triggers, in the case of Logistic Regression.

In a business context, this characteristic has concrete relevance, as it facilitates communication with business areas, auditability of the analytical process, and the incorporation of the model into marketing-oriented decision routines. Thus, the choice of the logistic model is not solely due to methodological preference, but to a combination of competitive performance, analytical parsimony, and greater transparency for applied use. This approach allows marketing teams to better understand the factors influencing “churn” and develop more effective and targeted retention strategies.

As an illustrative exercise of managerial application, the set associated with the confusion matrix of Logistic Regression was considered, totaling 196,525 instances. With 165,244 cases classified as positive by the model, obtained by the sum of true positives and false positives, a score-oriented strategy would reduce the volume of activations by approximately 16% compared to indiscriminate contact with the entire base. Assuming, for simulation purposes only, an average cost of R$ 5.00 per contact or retention action, the estimated cost would drop from R$ 982,625.00 to R$ 826,220.00.

This reduction corresponds to a potential saving of R$ 156,405.00 per cycle. This amount does not represent an actual financial gain observed by the company, as the simulation does not incorporate campaign conversion rates, customer value, incentive costs, or the incremental return of retention actions. Its objective is to demonstrate how prioritization based on “churn” risk can contribute to a more efficient allocation of the marketing budget. The ability to optimize marketing resources is a significant benefit for companies in the iGaming sector.

The results, therefore, indicate that the application of machine learning techniques to marketing analytics in iGaming is promising for anticipating “churn” within the 30-day horizon adopted in this study. The evidence obtained suggests that recent behavioral and transactional variables have a strong capacity to discriminate users with a higher probability of evasion and that different algorithms can offer distinct profiles of managerial utility. In this sense, the findings reinforce the viability of using predictive models to support segmentation, user prioritization, and the planning of retention actions in non-contractual environments.

In summary, “churn” prediction in iGaming can generate analytical and managerial value when integrated with marketing analytics practices, especially in segmentation, user prioritization, and retention planning activities. Among the evaluated models, conservative Logistic Regression was adopted as the main alternative because it combined competitive statistical performance, better behavior in AUC-ROC, AUC-PR, and “precision”, and greater interpretability for applied use, maintaining an “accuracy” of 0.8980 even in a more prudent specification. Random Forest, in turn, contributed as a relevant benchmark, broadened the understanding of the “trade-off” between coverage and selectivity, and reinforced the importance of recent activity and transactional frequency variables for reading evasion risk. Considered together, the results support the viability of using predictive models for the studied problem, but also indicate that their practical adoption should be accompanied by continuous validation, performance monitoring, and periodic review of the operational definition of “churn”.

4. Conclusion

This work aimed to develop and evaluate predictive models to estimate the probability of user churn on iGaming platforms, using behavioral and transactional variables, with the intention of supporting marketing analytics decisions in anticipating evasion. It was found that both Logistic Regression and Random Forest presented high predictive performance in the main churn scenario, defined by 30 days of inactivity. Logistic Regression stood out as the main alternative, combining high accuracy of 0.8980, better performance in AUC-ROC and AUC-PR, and greater precision, in addition to offering greater interpretability for managerial application. It was observed that Random Forest achieved slightly higher recall and equivalent F1-score, but with a higher cost in false positives. The findings indicated that the ability to discriminate users with a higher probability of evasion is strongly associated with recent activity and transactional frequency attributes, such as active days and number of deposits within 15 and 30-day windows. The main contribution of the study lies in demonstrating the viability of using machine learning techniques to generate analytical and managerial value in iGaming, supporting segmentation, user prioritization, and the planning of retention actions more efficiently, with the potential for marketing resource optimization.

However, the study presented some limitations that deserve consideration. The operational definition of churn adopted was specific to the analyzed context and dependent on a 30-day time window, which may not generalize to other scenarios. The predominance of very recent variables, although useful for short-term prediction, may not ensure the same level of predictive performance in longer horizons or on datasets with distinct structures. Furthermore, the evaluation of the models was carried out in a specific temporal scope, without additional validations in different periods or with external data. For future studies, it is suggested to perform tests with other inactivity horizons and classification thresholds, implement additional temporal validation strategies, calibrate predictive scores, and integrate with customer value metrics and the return on retention actions. It is also recommended to broaden the comparison between algorithms, incorporate new behavioral and promotional variables, and investigate the performance of the models in different operational contexts of the sector, in order to strengthen the analytical robustness and practical applicability of the obtained results.

Bibliographic References

Brasil. 2023. Lei nº 14.790, de 29 de dezembro de 2023. Dispõe sobre a modalidade lotérica denominada aposta de quota fixa, altera as Leis nº 5.768, de 20 de dezembro de 1971, nº 13.756, de 12 de dezembro de 2018, e n° 14.183, de 14 de julho de 2021, e dá outras providências. Brasília, DF: Presidência da República. Disponível em: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2023/lei/l14790.htm . Acesso em: 8 abr. 2026.

Brasil. Ministério da Fazenda. 2024. Regulamentação feita pela Secretaria de Prêmios e Apostas coloca Brasil em mercado regulado de apostas em 2025. Brasília, DF, 26 dez. 2024. Disponível em: https://www.gov.br/fazenda/pt-br/assuntos/noticias/2024/dezembro/regulamentacao-feita-pela-secretaria-de-premios-e-apostas-coloca-brasil-em-mercado-regulado-de-apostas-em-2025 . Acesso em: 8 abr. 2026.

Breiman, L. 2001. Random forests. Machine Learning, v. 45, n. 1, p. 5-32. DOI: 10.1023/A:1010933404324.

Erevelles, S.; Fukawa, N.; Swayne, L. 2016. Big data consumer analytics and the transformation of marketing. Journal of Business Research, v. 69, n. 2, p. 897-904. DOI: 10.1016/j.jbusres.2015.06.041.

Fader, P. S.; Hardie, B. G. S. 2007. How to project customer retention. Journal of Interactive Marketing, v. 21, n. 1, p. 76-90. DOI: 10.1002/dir.20074.

Gil, A. C. 2002. Como elaborar projetos de pesquisa. 4. ed. São Paulo: Atlas.

Jahromi, A. T.; Stakhovych, S.; Ewing, M. 2014. Managing B2B customer churn, retention and profitability. Industrial Marketing Management, v. 43, n. 7, p. 1258-1268. DOI: 10.1016/j.indmarman.2014.06.016.

James, G.; Witten, D.; Hastie, T.; Tibshirani, R. 2021. An introduction to statistical learning: with applications in R. 2. ed. New York: Springer. DOI: 10.1007/978-1-0716-1418-1.

Kaya, E.; Dong, X.; Pentland, A. S. 2018. Behavioral attributes and financial churn prediction. EPJ Data Science, v. 7, art. 41. DOI: 10.1140/epjds/s13688-018-0165-5.

Kumar, V.; Reinartz, W. 2018. Customer relationship management: concept, strategy, and tools. 3. ed. Berlin: Springer. DOI: 10.1007/978-3-662-55381-7.

Mena, G.; Coussement, K.; De Bock, K. W.; De Caigny, A.; Lessmann, S. 2024. Exploiting time-varying RFM measures for customer churn prediction with deep neural networks. Annals of Operations Research v. 339, p. 765-787. DOI: 10.1007/s10479-023-05259-9.

Merchie, F.; Ernst, D. 2022. Churn prediction in online gambling. Future Generation Computer Systems, v. 132, p. 108-123. DOI: 10.1016/j.future.2022.02.013.

Mitchell, T. M. 1997. Machine learning. ISBN 9780071154673. McGraw-Hill.

Prodanov, C. C.; Freitas, E. C. 2013. Metodologia do trabalho científico: métodos e técnicas da pesquisa e do trabalho acadêmico. 2. ed. Novo Hamburgo: Feevale.

Roberts, D. R.; Bahn

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