October 08, 2026
Customer conversion prediction as support for decision-making in fitness businesses
Customer Conversion Prediction as Support for Decision Making in Fitness Business
Leonardo da Silva Pianca; Thiago Gentil Ramires
DOI: 10.22167/2675-6528-202603139
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 study presented the development of a customer conversion predictive model for the fitness sector, employing Machine Learning techniques to support strategic decision-making. The database, composed of 7,696 records from a gym chain, underwent rigorous treatment and multicollinearity analysis (VIF). The Logistic Regression model demonstrated solid performance, achieving an accuracy of 58.93% and an area under the curve (AUC) of 62.5%. The results confirmed that operational and seasonal variables are robust predictors, with a highlight on registrations on Mondays, which showed a 157% higher conversion probability compared to other days. Complementarily, a benchmarking was performed with more complex algorithms; the Random Forest demonstrated greater effectiveness in capturing actual enrollments compared to the linear model, while XGBoost achieved maximum sensitivity (Recall of 68.54%). It was concluded that, although Logistic Regression is the ideal tool for extracting immediate commercial guidelines due to its analytical transparency, ensemble models revealed latent potential for future optimizations in capturing enrollments and reducing missed opportunities.
Keywords: Supervised analysis; Fitness; Odds ratio; Regression.
1. Introduction
The global fitness industry has shown continuous growth, evidenced by a total of approximately 174 million customers in over 201 thousand gyms in 2017, with an estimated financial turnover of 87.2 billion dollars. The International Health, Racquet & Sportsclub Association (IHRSA), as cited by ACAD Brasil (2018), set the goal of reaching 230 million practitioners globally by 2030. This scenario highlights the sector’s expansion potential and strategic relevance, making the understanding of factors influencing customer adherence and retention fundamental for gym competitiveness.
In Brazil, the fitness market has been consolidating as one of the most promising, with 15 million members enrolled in 2024 (FITNESS BRASIL, 2024) and an annual turnover exceeding 20 billion reais. This growth is driven by a 30% increase in the population’s exercise regularity over the past five years (ABJ; IBGE, 2026). However, the sector’s financial sustainability is challenged by customer retention. Data from the Panorama Setorial Fitness Brasil (2025) indicate that early dropout is a critical operational bottleneck, with the highest abandonment rate occurring within the first ninety days of enrollment. The lack of service differentiation and personalized attention exacerbate this phenomenon, increasing customer acquisition costs and demanding data-driven management strategies to identify profiles at risk of dropout (SEBRAE, 2024; FITNESS BRASIL, 2025).
Understanding the behavior of physical activity practitioners is complex and multifaceted. Studies on the motivation for practicing physical activity reveal that it changes throughout life. Santos et al. (2018) identified that the main motivation for weight training practitioners is the “health-fitness” factor, followed by “fun-interest”, “competence”, “appearance”, and “social”. Rodrigues (2022) adds that, in adolescence, adherence is more linked to intrinsic factors such as pleasure and group belonging, while in adulthood, health, disease prevention, and well-being become more relevant.
Brito et al. (2021) reinforce that motivators for weight training include body weight control, personal satisfaction, increased self-esteem, improved socialization, and reduction of cardiovascular diseases, stress, and depression. These elements demonstrate that the decision to join and remain in a gym is intrinsically linked to aspects of health, well-being, and social interaction. Customer loyalty, in turn, is a critical factor for the financial sustainability of gyms. Silva (2019) points out that the establishment’s location, the quality of service and instructors, the available infrastructure and equipment, as well as general satisfaction and price, are recurring factors for student retention. These aspects indicate that tangible and intangible variables exert significant influence on customer retention.
Given the complexity of factors influencing customer adherence and retention in the fitness sector, and the need for more assertive strategies to mitigate early churn, the use of advanced analytical tools becomes imperative. The application of Machine Learning techniques emerges as a promising solution to transform operational data into business intelligence, enabling the identification of patterns and the prediction of conversion behaviors. Thus, this study is justified by the pursuit of optimizing strategic decision-making in fitness businesses, and aims to identify the main factors influencing the conversion of leads in a gym chain, transforming operational data into business intelligence.
2. Material and Methods
The research was quantitative and exploratory in nature, with a methodological approach based on Machine Learning techniques. The study focused on the development of a predictive customer conversion model, aiming to transform operational data into business intelligence for a gym chain. The unit of analysis consisted of visit records of potential customers.
Data collection was carried out from a secondary internal database, extracted from the CRM system of a gym chain. This chain operates with three units located in the city of Rio de Janeiro, RJ, and serves approximately 4,000 students. The analyzed database comprised 7,696 customer visit records, collected from January to December 2025.
To enable the application of Machine Learning algorithms, the raw data were submitted to a cleaning and transformation process. Initially, the identities of the collaborators responsible for the registrations were pseudonymized, using the technique of preserving the name followed by the initial of the last surname, ensuring anonymity and the protection of sensitive data.
The “How did you hear about us” variable was treated to remove structural noise, such as special characters and blank spaces. Missing data in the “How did you hear about us” and “Registered by” variables were filled with the categorical label “Nao_Informado”, representing 1.2% of observations, to avoid sample reduction and allow evaluation of their correlation with the conversion rate.
The dependent variable, named “target”, was constructed in a binary form from the “Customer ID” column. Records with a customer identification code were classified as 1 (Converted), indicating the enrollment was completed, while records with null fields were classified as 0 (Not Converted).
The original “Cadastro” column, in datetime format, was split to create the “mes_cadastro” and “dia_semana_cadastro” variables, extracting the monthly component and the corresponding day of the week (integers from zero to six). Subsequently, these variables were converted to categorical types, using three-character abbreviations, and subjected to the One-Hot Encoding technique to treat each month and day as independent events.
The “Shift” variable was created to investigate the influence of the service period on lead conversion. As the CRM system did not record the exact contact time, this variable was derived by mapping consultants, based on the unit’s fixed work schedule. Morning consultants were labeled as ‘Morning’ and afternoon/night consultants as ‘Night’.
Gender inference was performed using the `gender_guesser.detector` library, which classifies proper names into ‘Male’, ‘Female’, or ‘Unknown’. Names classified as “Unknown” were manually analyzed and corrected with an external mapping dictionary, ensuring the completeness of the sociodemographic categorical variable.
The variable “Age”, absent in the raw CRM records, was synthetically generated through normal (Gaussian) distribution simulation, with a mean of 32 years and a standard deviation of 12 years, based on empirical knowledge of the gym’s demographic profile. Clipping technique was applied to limit the values to the biological and operational interval of the unit (14 to 85 years).
To facilitate model interpretation and reduce variability, continuous age was discretized into four strategic categories, resulting in the creation of the “Age_Range” column: “Young Audience / Student” (14-25 years), “Adults / Economically Active Audience” (26-40 years), “Mature Audience” (41-60 years), and “Senior Audience” (61+ years).
Categories with low sample density in the “How did you hear about us?” variable were grouped. Observations with representativeness below 1% were consolidated into “Others”, including similar content such as “Passing”, “Facade”, and “Invitation”, to focus statistical weight on the most relevant commercial processes and avoid distortions in the results.
Similarly, in the “Registered by” variable, records attributed to temporary consultants, technical support, or occasional collaborators with very low frequency (less than 45 observations each) were consolidated into the ‘Others/Occasional’ group. This measure aimed to mitigate the impact of categories with high variance and few occurrences.
To select the most relevant variables and ensure the feasibility of the analysis, the Chi-Squared (X²) independence test was applied with the aid of the `scipy.stats` library (Fávero and Belfiore, 2024). The nominal qualitative variables were subjected to the dummification process (dummy encoding) using the `get_dummies` function from the `pandas` library, with the `drop_first=True` parameter to remove a reference category.
Multicollinearity analysis was performed by calculating the Variance Inflation Factor (VIF) using the `statsmodels` library (Pedregosa et al., 2011). The procedure consisted of applying the `variance_inflation_factor` function to the matrix of predictor variables, previously converted into dummy variables.
The predictive modeling strategy focused on applying Binary Logistic Regression as the main model, due to its statistical robustness and high interpretability (Fávero and Belfiore, 2024). This model was used to estimate the probability of a dichotomous event occurring (lead conversion or non-conversion) through the logit function. The practical implementation was carried out with the `Scikit-learn` library (Pedregosa et al., 2011).
In order to validate the results and seek accuracy gains, the Random Forest and XGBoost algorithms were implemented for a comparative approach. Random Forest, based on ensemble learning, creates multiple decision trees independently. XGBoost, a Gradient Boosting algorithm, builds trees sequentially to correct prediction errors (Chen; Guestrin, 2016). Both were implemented with the `RandomForestClassifier` and `XGBClassifier` classes from their respective libraries.
For the evaluation of the performance of predictive models, binary classification metrics were used: accuracy, area under the ROC curve (AUC), and confusion matrix. Accuracy measured the proportion of correct predictions. AUC assessed the model’s discriminatory capacity. The confusion matrix allowed for a detailed analysis of performance, identifying true positives, true negatives, false positives, and false negatives, with computational implementation via the `scikit-learn` library (Pedregosa et al., 2011).
3. Results and Discussion
The data analysis began with the application of the Chi-Square test (X²) to verify the association between the explanatory variables and the dependent variable of lead conversion. The results indicated that, with the exception of the variables ‘Shift’ and ‘Gender’, all other variables showed a statistically significant association with lead conversion, with p-values below 0.05. This finding allowed the rejection of the null hypothesis of independence, confirming that the database possessed the necessary conditions for the construction of a logistic regression model with statistically supported patterns. The initial validation of variables is crucial to ensure that the predictive model is based on genuine relationships and not on random associations, as recommended by Fávero and Belfiore (2024).
Complementarily, multicollinearity analysis was performed using the Variance Inflation Factor (VIF) to assess the correlation between independent variables. It was observed that the ‘Shift’ variable presented an excessively high VIF, indicating a direct linear dependence on the ‘Registered by’ variable. This phenomenon occurred because the service shift was a direct derivation of the consultants’ fixed work schedule, making the information contained in ‘Shift’ redundant concerning the employees’ names. The removal of variables with high multicollinearity is essential to ensure the stability of regression coefficients and the statistical validity of the model, avoiding distortions in the interpretation of the individual impact of each predictor.
The characterization of the target variable, lead conversion, revealed a rate of 56.9% of effective enrollments (4,382 records) in contrast to 43.1% of leads that did not convert (3,314 records). This scenario, unlike many commercial contexts where conversion is a rare event, demonstrated a robust volume of examples for both classes. To mitigate any bias in predictions due to the slight disparity between classes, the strategy of balanced weights was maintained in the logistic regression model. This procedure ensured that the characteristics of non-converted leads were weighted with the same relevance as those of enrolled students, providing greater impartiality and scientific rigor to the probability estimates.
Performance of the Logistic Regression Model exclusively with operational and temporal variables
The Binary Logistic Regression model was trained to predict the probability of conversion of new students, using data collected between January and December 2025. Initially, the model presented an accuracy of 53% and an Area Under the Curve (AUC) of 56%. These preliminary results indicated the need for refinement in feature engineering, as the excess of variables and categories with low sampling was introducing noise into the algorithm’s learning. The optimization of predictive capacity was then sought through structural adjustments in the database, aiming to improve the model’s performance and interpretability.
The database refinement involved the suppression of the ‘Gender’ variable, which showed low statistical relevance, and the ‘Shift’ variable, which presented a high correlation with ‘How did you hear about us’. Additionally, continuous age was replaced by age group clusters, allowing for a better capture of the specific motivations of each life cycle, as substantiated by Rodrigues (2022). The exclusion of records categorized as ‘Others’, referring to channels and consultants with fewer than 45 observations (2.2% and 1.5% of the total), eliminated operational outliers and distortions caused by atypical behaviors, which generated unrealistic conversion rates, such as 0% or 100%, due to insufficient sampling.
After the adjustments, the accuracy of the Logistic Regression model reached 58%, and the AUC reached 63%. Accuracy represents the proportion of correct classifications over the total test sample, while AUC suggests that the model has a discrimination capacity superior to chance, although still with room for optimization. The confusion matrix revealed 722 correctly classified conversions (True Positives) and 625 non-conversions (True Negatives). However, the volume of errors, with 593 False Negatives and 369 False Positives, reflects the complexity of behavioral variables in the fitness sector, highlighting the difficulty in predicting with total precision all cases of conversion and non-conversion.
An additional refinement was made by excluding the ‘Others/occasional’ category from the ‘Registered by’ variable, which grouped records from temporary consultants or those with very low registration frequency. This adjustment aimed to eliminate random behaviors and focus the model on the more standardized and representative commercial processes of the academy. The objective was to enhance the practical utility of the model by concentrating the analysis on data generated by specialized sales consultants, which aligns the algorithm with the academy’s actual commercial process and proves the importance of standardizing service for conversion predictability.
This refinement resulted in a direct improvement in the algorithm’s ability to identify real conversions, with True Positives increasing from 722 to 745. In parallel, False Negatives were reduced from 593 to 552. This increase in the Recall (sensitivity) metric is fundamental for the fitness business, as the cost of not identifying a potential client with a high probability of enrollment (False Negative) is higher than the cost of approaching a lead that does not convert. The reduction in omission error allows the sales team to focus its efforts on a broader and more qualified lead base, capturing opportunities that would be ignored due to the noise caused by non-standardized records.
Chance analysis
The use of Logistic Regression allowed for the calculation of the Odds Ratio, an essential metric for quantifying the impact of each variable on the probability of enrollment closure. The results revealed that the Curicica Unit presented the highest odds ratio, with a conversion probability 157.3% higher than other units. This data suggests that specific management factors or customer profiles of this unit contribute significantly to enrollment success, indicating superior performance that can serve as a benchmark for other units in the network.
The service provided by consultant Giovanna F. also stood out, increasing the conversion chance by 76.1%. This finding highlights the importance of individual sales consultants’ performance and the quality of service in the lead conversion process. The performance of specific consultants can be an indicator of good sales and customer relationship practices, which can be replicated to improve the effectiveness of the team as a whole.
Seasonal and operational variables proved to be robust predictors of conversion. Registrations made on Mondays showed a 109.2% higher conversion probability compared to other days of the week, while registrations in January showed an 82.5% greater chance. The acquisition channel “Facade (Walked by)” also stood out, with a 69.6% higher conversion probability. These results confirm that operational and seasonal variables are robust predictors, with a particular emphasis on registrations on Mondays, which showed a 157% higher conversion probability compared to other days, as per the approved summary.
Analysis of the impact of synthetic variables (Age and Gender)
To enrich the analysis, an experiment was conducted including the synthetic variables Age (age groups) and Gender. Although the inclusion of this demographic data did not result in a significant gain in overall performance, maintaining Accuracy at 59.23% and AUC at 0.6254, values very close to those obtained by the strictly operational model, the interpretation of the coefficients offered relevant qualitative insights that corroborate the evidence from Santos et al. (2018) and Rodrigues (2022) regarding the change in motivation with advancing age.
In the expanded model with synthetic variables, the senior group (61+ years) presented itself as a prominent predictor, with an Odds Ratio of 1.99, indicating that this audience has almost double the chance of conversion compared to other age groups. This finding validates the theory that the mature audience has a greater focus on health and well-being, presenting a more objective decision cycle. However, when comparing the two scenarios, it was observed that the suppression of synthetic variables did not compromise the predictive logic of the model.
Without age and gender data, variables related to physical location and individualized consultant service took statistical prominence. The stability of the metrics (AUC of 0.625 in the operational model versus 0.630 in the synthetic model) suggests that operational data already implicitly carries the behavior of different demographic profiles. Therefore, for the operational reality of the academy, the model focused on real CRM system variables proved to be more robust and less dependent on demographic assumptions, allowing for strategic decision-making based on auditable and directly observable facts.
Comparative Evaluation of models
In order to validate the effectiveness of the solution and evaluate the performance of different machine learning architectures, a comparative test was carried out between Logistic Regression, Random Forest, and XGBoost. The main objective of this stage was to verify if models with higher computational complexity (non-linear) would deliver better predictive effectiveness. All models were subjected to the same preprocessing, using a 70% data split for training and 30% for testing, with stratification applied to maintain the proportion of the target variable in both sets.
The results obtained demonstrated that Logistic Regression achieved a solid overall performance, reaching an accuracy of 58.93% and an AUC-ROC of 0.625. This performance consolidated it as the most balanced solution for the academy’s operational scenario, mainly due to its high interpretability, which facilitates the extraction of immediate commercial guidelines. The analytical transparency of Logistic Regression allows managers to easily understand which factors are driving conversion, aiding in the formulation of targeted strategies.
Upon deepening the analysis through the confusion matrix, it was observed that non-linear models, such as Random Forest and XGBoost, managed to capture conversion patterns that logistic regression did not achieve. Random Forest, for example, recorded 532 False Negatives, compared to 552 for the linear model, and correctly identified 765 enrollments. This result indicates that Random Forest was more effective in capturing actual enrollments compared to the linear model, as per the approved summary.
The XGBoost algorithm presented the highest test sensitivity, with a Recall of 68.54%, correctly identifying 889 enrollments and registering only 408 False Negatives. These numbers reveal that, despite Logistic Regression being the most accurate tool for current diagnosis and for identifying business levers, there is latent potential in more complex algorithms. Ensemble models, such as Random Forest and XGBoost, demonstrated a superior ability to identify leads that would effectively convert, significantly reducing missed opportunities.
The results obtained with ensemble models serve as a benchmark for superior performance, indicating that delving into more robust machine learning techniques constitutes a promising path for future optimizations of the network’s conversion rate. Although Logistic Regression is ideal for extracting immediate business guidelines due to its analytical transparency, ensemble models reveal latent potential for future optimizations in capturing enrollments and reducing missed opportunities, as highlighted in the study’s abstract. The ability to identify a higher volume of True Positives and reduce False Negatives is crucial for the sustainability and growth of the fitness business.
4. Conclusion
The present study sought to identify the main factors influencing lead conversion in a gym chain, transforming operational data into business intelligence. It was found that the Logistic Regression model, after rigorous data treatment and refinement, demonstrated solid performance, achieving an accuracy of 58.93% and an area under the curve (AUC) of 62.5%. The findings confirmed that operational and seasonal variables are robust predictors, with a highlight on registrations made on Mondays, which showed a 157% higher conversion probability compared to other days of the week. Additionally, the Curicica unit and the performance of specific consultants, such as Giovanna F., were identified as factors that significantly increase enrollment chances. The main practical contribution of this work lies in the Logistic Regression’s ability to provide immediate and transparent commercial guidelines, allowing managers to optimize prospecting and customer service strategies based on auditable data.
Despite the robustness of the operational model, the complexity of behavioral variables in the fitness sector imposed challenges in prediction, resulting in a volume of False Negatives and False Positives that reflects the inherent difficulty in predicting with total accuracy all conversion cases. One identified limitation was the absence of detailed demographic data in the original CRM, which led to the generation of synthetic variables for experimentation, although the model focused on operational data proved more robust. For future studies, benchmarking with more complex algorithms, such as Random Forest and XGBoost, revealed a latent potential for optimizations in capturing enrollments and reducing missed opportunities. It is suggested to deepen the refinement of XGBoost, using explainability techniques, and to improve the collection of prospecting data, including information such as ZIP code, age, and physical activity history, in addition to incorporating churn prediction models for more complete customer lifecycle management.
Bibliographic References
ABJ; IBGE, 2026 [Referência completa não encontrada no documento original]
ACAD Brasil. (2018). Mercado mundial do fitness: principais players e mudanças no top ten. Revista ACAD Brasil, Ano 20(3), n. 82. Disponível em: https://acadbrasil.com.br/wp-content/uploads/2019/03/edicao-82.pdf. Acesso em: 19 set. 2025
Brito11, Geisson C., et al. “Adesão a prática de musculação nas academias do Brasil.” Revista CPAQV-Centro de Pesquisas Avançadas em Qualidade de Vida| Vol 13.1 (2021): 2.
CHEN, Tianqi; GUESTRIN, Carlos. XGBoost: A scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: ACM, 2016. p. 785–794.
FITNESS BRASIL. Panorama Setorial Fitness Brasil 2025. 4. ed. São Paulo: Fitness Brasil, 2025. Disponível em: https://www.fitnessbrasil.com.br. Acesso em: 8 abr. 2026
FÁVERO, Luiz Paulo; BELFIORE, Patrícia. Manual de análise de dados: estatística e machine learning com Excel®, SPSS®, Stata®, R® e Python®. 2. ed. Rio de Janeiro: GEN LTC, 2024.
PEDREGOSA, Fabian et al. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, v. 12, p. 2825–2830, 2011.
Rodrigues, Giovanna Oliveira. “Motivação para a prática de atividade física: uma revisão de literatura.” (2022).
Santos, Francielle Carvalho, et al. “Motivação a prática de atividades físicas: um estudo com praticantes de musculação.” Revista Científica UNIFAGOC-Multidisciplinar 3.1 (2018).
SEBRAE. Panorama Mercado De Academias No Brasil: Dados E Tendências. Curitiba: Sebrae/PR, 2024. Disponível em: https://sebraepr.com.br. Acesso em: 8 abr. 2026
Silva, Luis Rafael Cardoso da. “Desafios da fidelização de clientes em academias de ginástica: uma revisão da literatura.” (2019).
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