February 03, 2026
Logistic Regression for Production Volume Estimation in Vehicle Manufacturers
Caio Guilherme Rezende Silva; Elisa Morandé Sales
Summary prepared by the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege focused on synthesis and writing.
This work develops a decision support tool to assist analysts in production control and future volume forecasting, aiming to reduce manual effort and increase information reliability based on statistical methods. The tool estimates the volume of vehicles to be produced and made available for invoicing, calculating the probability of each unit being delivered within the current month. The system also provides forecasts segmented by market, model, and version, essential information to support sales planning, logistics, and marketing, aligning production capacity with the organization’s commercial goals.
The solution fits into the context of the competitive Latin American automotive sector; Brazil is the main consumer (Mordor Intelligence, 2024). With more than fifty brands competing in the passenger car and light commercial vehicle market, agility and precision in strategic planning are crucial differentiators (Sindicato dos Metalúrgicos, 2023). Automakers face the challenge of adapting their sales plans to maintain or expand their market share. To do this, they need a clear and forward-looking view of their product availability.
One of the most critical pieces of information to obtain with precision is the exact volume of vehicles available for invoicing on a specific date, detailed by model, version, and country. The difficulty lies in the multiple variables that influence the production process, such as assembly line pace, programming date, component complexity, quality verification processes, and potential delays. The uncertainty generated by these factors impacts the company’s ability to commit to deadlines and optimize commercial strategies, which in 2024 registered 536,604 light commercial vehicles registered in Brazil (G1, 2025).
Given this scenario, the study proposes to answer what the productive volume available for sale will be on a future date, using machine learning techniques. The approach analyzes a set of variables from the production cycle to predict the probability of each unit being available for invoicing before the month’s accounting closing. The choice of binary logistic regression as a model is justified by its effectiveness in classification problems, being a robust and consolidated method in areas such as the financial sector for credit risk analysis, calculating the probability of an event occurring based on observed characteristics (Araújo & Carmona, 2007).
The application of a predictive model represents a significant evolution compared to traditional methods, based on the subjective experience of analysts and complex spreadsheets. By automating and grounding the analysis in a statistical model, the tool increases the accuracy of predictions and frees up professionals to focus on higher value-added activities. The ability to simulate different production scenarios, adjusting variables such as the production plan, allows for more proactive and adaptable management, aligned with the dynamics of a market that demands quick and well-informed decisions.
The method employed was binary logistic regression, a supervised learning technique from the field of machine learning (Blum, n.d.). The choice is justified by the nature of the response variable: determining whether a vehicle will be delivered within the month (event) or not (non-event), a dichotomous classification. Supervised learning uses historical labeled data to train an algorithm, allowing it to identify patterns and mathematical relationships between variables (Monard & Baranauskas, n.d.). Once trained, the model applies this knowledge to new data to predict the outcome. Binary logistic regression classifies each individual based on a probability, ranging from 0 to 1, of the event of interest occurring (Gonçalves et al., 2013).
The database was extracted from the production control systems, comprising the last three months of production and the plan for the current month. This period was defined to train the model with recent data, reflecting current operational patterns. The variables collected for each vehicle included the production scheduling date, the time (in days) in the yard for quality checks, and characteristics such as model, engine, transmission, and destination country. The dependent variable was the binary information indicating whether the vehicle was produced and made available for sale in the same month. The Python programming language was adopted for data processing and algorithm development, due to its versatility and libraries for data analysis and machine learning (Virtanen & Gommers, 2019).
The data preprocessing was a fundamental step. The variable “Days to closing” was calculated dynamically, depending on the vehicle’s status: for cars already delivered, it was the difference between the last business day of the month and the availability date; for cars under review, the difference between the last business day and the analysis date; and for cars with future production, the difference between the last business day and the scheduled date. Qualitative variables, such as model and engine, were transformed into boolean (dummy) variables. The database was divided into a training set (historical data) and a prediction set (vehicles in process or future). The mathematical formulation of the model follows the logistic probability equation, which transforms a linear combination of variables into a probability between 0 and 1 (Fávero & Belfiore, 2023).
For the implementation, the Python statsmodels library was used, with the glm (Generalized Linear Models) function (Statsmodels, 2024). The model was trained with the retroactive database, generating coefficients that represent the influence of each variable. The quality of the fit was evaluated by metrics such as Log-Likelihood and Pseudo R² (Aprender estática fácil, 2025; Psicometria. online, 2025). The statistical significance of each variable was verified through the P-value, considering those with a value below 0.05 as relevant (Medium, 2024). After training, the model was validated with a confusion matrix for the calculation of overall efficiency (EGM), sensitivity, and specificity, in addition to the ROC curve analysis to assess classification performance.
The results obtained demonstrated the model’s high predictive capacity. During a three-month testing period (February, March, and April), forecasts were compared with actual results at the end of each month. The consolidated analysis revealed an average accuracy of 97.5%, indicating an average discrepancy of only 2.5% between the forecasted and the effectively realized volume. This level of precision represents a substantial advancement compared to previous methodologies, providing the sales and supply chain team with a more reliable database for planning.
The detailed analysis of the parameters generated by the statistical model confirmed the relevance of the variables. “Days in quality” and “Days to closing” presented the lowest P-values and the highest absolute values in the Z statistic, confirming them as the predictors with the greatest impact. Graphical investigation revealed an inversely proportional correlation between days in quality control and the probability of delivery. On the other hand, the variable “Days to closing” showed a positive correlation, indicating that vehicles scheduled for the beginning of the month have a higher probability of being finalized on time.
A crucial aspect was the definition of the cutoff point, the probability threshold for classifying a vehicle as “delivered”. Tests were performed with different cutoffs, from 0.4 to 0.9, and for each, the overall model efficiency (EGM) was calculated. The analysis demonstrated that the maximum EGM of 93.5% was achieved with a cutoff of 0.55 on the training set. The analysis of the sensitivity and specificity curve, along with the ROC curve, corroborated the model’s robustness. For the final application, a cutoff of 0.7 was adopted, a decision that balances maximizing correct predictions with a more conservative stance, aligned with the company’s strategy.
The model’s performance by market and vehicle, detailed in Table 7 of the original study, showed exceptionally accurate results for the Brazilian market, with variations oscillating between -3.1% and 0.6%. For export markets such as Argentina, Chile, and Colombia, although the overall accuracy remained high, a slightly greater variability was observed. For example, in February, the prediction for Model A in Argentina had a deviation of 12.9%, while in April, for Model A in Colombia, the deviation was 11.2%. These isolated discrepancies did not compromise the aggregate result, which presented a total delta of only 1.4% in February, 0.2% in March, and 0.7% in April, demonstrating the model’s consistency.
Beyond accuracy, one of the most significant gains was the reduction in analysis time. The process, which previously demanded an average of five hours of manual work, now takes approximately two minutes to execute. This time optimization increased productivity and allowed analyses to shift from weekly to daily, offering an always up-to-date view of production status and enabling faster reactions to deviations from the plan.
The tool transcended its forecasting function and consolidated itself as an instrument for scenario building. The ability to adjust parameters, such as the volume of future programming, and obtain a new delivery estimate in real-time, became valuable. Management began to use the model to simulate the impact of measures such as the implementation of overtime or Saturday work shifts, quantifying the potential gain in billing volume and allowing for a more precise cost-benefit analysis.
The success of the implementation in the pilot plant motivated the expansion of the project. The model was adapted and implemented in three other plants of the automotive group, with the objective of standardizing the volume forecasting process and disseminating a culture of data-driven management. This expansion reinforces the scalability of the solution and its potential to replace intuition-based analyses with a systematic, transparent, and statistically robust process.
The discussion of the results shows that the application of machine learning, through logistic regression, offers a powerful solution to a critical challenge in the automotive industry. The model’s ability to learn from historical data and identify subtle patterns allows it to generate predictions that surpass the accuracy of manual methods. The developed tool is not just a predictor of volumes, but a decision support system that provides insights into production bottlenecks and the impact of operational decisions, enabling the company to navigate a competitive market with greater confidence.
It is concluded that the objective was achieved: it was demonstrated that the application of a binary logistic regression model is a viable and high-impact tool to accurately estimate the production volume available for billing in vehicle manufacturers, subsidizing strategic decision-making based on data. The successful implementation validates the methodology as an effective solution for optimizing sales, logistics, and marketing planning. The quantitative benefits were significant, notably the increase in prediction accuracy to an average of 97.5% and the reduction in analysis time from five hours to two minutes. This efficiency allowed for more dynamic management, transforming the tool into a pillar for scenario building and strategic decision-making, such as evaluating the return on investment in overtime. The project’s success culminated in its expansion to three other plants in the group, consolidating a data-driven management approach and reinforcing the applicability of machine learning techniques to solve complex problems and generate tangible value in the automotive sector.
References:
Estatística Fácil. 2025. What is: Log-Likelihood. Available at: https://estatisticafacil. org/glossario/o-que-e-log-likelihood/. Accessed on: June 14, 2025.
Araújo, E.; Carmona, C. 2007. Development of Credit Scoring Models with Logistic Regression Approach for the Default Management of a Microcredit Institution. Revista Contabilidade Vista & Revista: 107-131.
Blum, A. n. d. Machine Learning Theory. Department of Computer Science, Carnegie Mellon University… p. 1-4.
Fávero, L. Belfiore, P. 2023. Manual of data analysis. 2nd ed. GEN. RJ, Brazil.
G1. New vehicle sales in Brazil increased by 14.1% in 2024.2025. Available at: https://g1. globo. com/carros/noticia/2025/01/08/venda-de-veiculos-novos-no-brasil-tem-alta-de-141percent-em-2024. ghtml. Accessed on: March 28, 2025.
Gonçalves, E.; Gouvêa, M.; Mantovani, D. 2013. Credit risk analysis with the use of logistic regression. Revista Contemporânea de Contabilidade: 140-160.
Medium. 2024. Inference in logistic regression. Available at: https://medium. com/@paulinhomagno/infer%C3%AAncia-na-regress%C3%A3o-log%C3%ADstica-792475201c38. Accessed on: June 15, 2025.
Monard, M. Baranauskas, J. n. d. Concepts on Machine Learning.
Mordor Intelligence. 2024. Latin America passenger car market size. Available at: https://www. mordorintelligence. com/pt/industry-reports/latin-america-passenger-car-market-outlook. Accessed on: March 28, 2025.
Psicometria. online. 2025. Logistic regression: pseudo R². Available at: https://www. blog. psicometriaonline. com. br/regressao-logistica-pseudo-r2/. Accessed on: June 14, 2025.
Metalworkers’ Union. 2023. Brazil already has 50 car brands. Available at: https://smabc. org. br/brasil-ja-tem-50-marcas-de-carros/. Accessed on: March 28, 2025.
Statsmodels. 2024. Statistical models, hypothesis tests, and data exploration. Available at: https://www. statsmodels. org/stable/index. html. Accessed on: June 14, 2025.
Virtanen, P. Gommers, R., J. 2019. Fundamental Algorithms for Scientific Computing in Python, University of Jyväskylä.
Executive summary from the Final Course Work of the Specialization in Data Science and Analytics of the MBA USP/Esalq
Learn more about the course; click here:
