February 02, 2026
Data science for tire sales forecasting in the Brazilian market
Dalton Gonçalves Silva dos Reis; Elisa Morandé Sales
Summary prepared by the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege focused on synthesis and writing.
The objective of this work was to develop and validate a predictive model for the sale of passenger tires in Brazil, segmented between the original equipment manufacturer (OEM) and replacement markets, aiming to provide a strategic tool for production planning and inventory management. The research sought to identify the most influential variables in each segment and apply time series models to generate assertive forecasts, optimizing resources and reducing costs. The need for such a model arises from the complexity of the tire production chain; inaccurate demand forecasting can lead to imbalances such as high storage costs due to excess inputs or production disruptions due to demand underestimation.
The relevance of the study is anchored in the magnitude of the Brazilian tire market, which moves more than 50 million units annually and is influenced by factors such as the performance of the automotive industry and macroeconomic indicators. Data science is fundamental for transforming raw data into actionable knowledge. According to Grus (2019), data analysis allows for the extraction of hidden patterns. This process of transforming data into information and knowledge constitutes what Fávero and Belfiore (2024) describe as the hierarchy of knowledge, a pillar for strategic decision-making and for consolidating a competitive advantage.
The approach focuses on robust and interpretable statistical models such as SARIMA and SARIMAX. The literature, such as Taylor’s (2015) study which used artificial neural networks (ANN) and achieved a coefficient of determination (R-squared) of 66%, serves as a reference. This study seeks to advance by testing alternative models and incorporating exogenous variables from the Brazilian context to obtain superior predictive performance, allowing companies in the supply chain to adjust their production plans and optimize logistics.
The application of techniques such as time series analysis represents an evolution in management, allowing for a proactive rather than reactive stance. VanderPlas (2016) highlights that data science empowers professionals to solve concrete problems, such as uncertainty in tire demand. By providing a reliable predictive model, the study offers practical support for more informed decisions, aligning supply and demand. The research is justified by its direct applicability in the industry, with the expectation that the developed models will generate a measurable impact on operational efficiency and reinforce the value of data analysis as a strategic asset for industrial competitiveness.
The methodology is a quantitative and applied research, based on the analysis of time series with public and structured data. The target variable, sales of passenger car tires, was obtained from monthly reports of the National Association of the Tire Industry (ANIP) from January 2018 to December 2024, totaling 84 observations. This period was chosen to capture different economic cycles, including the COVID-19 pandemic. To enrich the models, exogenous variables were incorporated: vehicle production and sales (ANFAVEA), sectoral indicators for rubber products (CNI), tire imports (ComexStat), and economic indicators such as consumer default (Serasa). According to Joseph and Tackes (2024), the incorporation of correlated variables builds more robust models. The time series used are stochastic in nature, with future values partially random, but describable by patterns.
Processing and modeling were performed in Python, using libraries such as Pandas and Numpy for data manipulation, and Matplotlib and Seaborn for visualization. Statistical modeling was implemented with the statsmodels library. The seasonaldecompose function was used to decompose the series into trend, seasonality, and residual (Oliveira et al., 2024). Stationarity, a prerequisite for ARIMA models, was verified with the Augmented Dickey-Fuller test (adfuller), which rejects the null hypothesis of non-stationarity for p-values below 0.05. For the selection of parameters (p, d, q) and (P, D, Q), the autoarima function from the pmdarima package was used, which selects the configuration that minimizes the Akaike Information Criterion (AIC), a metric that balances fit and parsimony (Hyndman and Athanasopoulos, 2021). The central models were SARIMA and SARIMAX, which extends SARIMA by including exogenous variables. Performance was evaluated by the Mean Absolute Percentage Error (MAPE), calculated with the function from the sklearn package. The series of 84 observations was divided into a training set with 72 observations (January 2018 to December 2023) and a test set with 12 observations (January to December 2024).
The analysis of the sales segment for original equipment manufacturers (OEMs) revealed an abrupt drop in 2020 due to the pandemic and a strong component of annual seasonality. The Augmented Dickey-Fuller test confirmed the stationarity of the series (p-value = 0.000896), not requiring differentiation (d=0). The auto_arima tool suggested a SARIMA(2, 0, 0)(0, 1, 1) model, which, after training and validation, resulted in a MAPE of 8.74%, equivalent to an accuracy of 91.26%.
To improve the model, the correlation with the production of passenger vehicles was investigated, which proved to be strong and positive (0.9). The inclusion of this exogenous variable in the model, transforming it into a SARIMAX, maintained the same order of ARIMA(2, 0, 0)(0, 1, 1) parameters. When training and validating the new model, the obtained MAPE was 7.92%, improving assertiveness to 92.08%. This reduction in error confirms the predictive value of the variable. The comparison of forecasts for the first four months of 2025 with actual data showed the superiority of SARIMAX, with an accumulated difference of 210,139 units against 218,523 from the SARIMA model.
The analysis was replicated for the aftermarket. The time series also proved to be stationary. The auto_arima identified the SARIMA(1, 0, 0)(1, 1, 1) model as the most suitable, which after training and validation presented a MAPE of 12.38% (assertiveness of 87.62%). The error being higher than that of the assembly line market is expected, as the aftermarket is influenced by a wider range of random factors linked to the final consumer’s decision, as pointed out by Morettin (2018).
For the SARIMAX model of the aftermarket, correlation analysis revealed a moderate negative correlation (-0.5) between tire sales and the Serasa default index. The logic is that an increase in indebtedness leads to the postponement of expenses such as tire replacement. Incorporating the default index as an exogenous variable, the SARIMAX model, with the same ARIMA(1, 0, 0)(1, 1, 1) order, was trained and validated. The result was a significant improvement in performance, with MAPE dropping to 9.71% (assertiveness of 90.29%). The forecast for the first four months of 2025 with this model also proved to be more accurate, with an accumulated difference of 475,660 units compared to the actual, versus 687,898 from the SARIMA model.
The results for the car manufacturer market indicated that the model with two variables (Tire Sales, New Vehicle Production) presented the lowest AIC and the lowest MAPE (7.92%), being the most efficient option. The addition of a third variable (IBGE Industry Index) did not bring improvements, suggesting information redundancy. For the replacement market, the model with two variables (Tire Sales, Default Rate Index) achieved the lowest MAPE (9.71%), confirming itself as the best choice. Although its AIC was slightly higher than that of the univariate model, the substantial improvement in MAPE justified the choice of SARIMAX. The attempt to add a third variable (Vehicle Trade Activity) resulted in a worsening of the MAPE to 14.33%, validating that not every correlated variable contributes positively to prediction, a principle in feature selection discussed by Nielsen (2021).
The results demonstrate that the application of the SARIMAX model is highly effective for forecasting tire sales in Brazil. For the automotive manufacturers market, the inclusion of vehicle production was crucial to achieve an accuracy greater than 92%. For the replacement market, the incorporation of the consumer default index allowed for an increase in accuracy to over 90%. The discussion of the results reinforces the distinction between the two markets. The automotive manufacturers segment operates in a more deterministic manner, with a direct relationship to vehicle production. In contrast, the replacement market is more stochastic, subject to fluctuations in consumer confidence and economic health. The research validates the hypothesis that customized models for each segment, using appropriate exogenous variables, outperform univariate models. The accuracy achieved, in both cases above 90%, represents an advancement compared to previous benchmarks, such as Taylor’s study (2015), and offers a tool of great practical value.
The study successfully demonstrated the viability and effectiveness of applying SARIMA and SARIMAX time series models for forecasting passenger tire sales in the Brazilian market. The developed models achieved accuracy rates exceeding 90% for both segments analyzed, validating them as robust tools for supporting production planning and inventory management. The research identified that the incorporation of specific exogenous variables for each market, such as vehicle production for the OEM segment and the default rate index for the replacement segment, was fundamental to improving forecast accuracy. The practical implications of these results are direct, offering industries a quantitative method to anticipate demand and optimize resource allocation.
Despite the positive results, the study has limitations. The main one lies in the forecast horizon; the models are best suited for short-term forecasts, ideally three to four months, as the uncertainty of variables increases over longer horizons. Continuous application with monthly adjustments is recommended. As a suggestion for future work, exploring machine learning models and neural networks such as LSTM or GRU could bring improvements. Furthermore, expanding the analysis to include other tire categories, such as truck and bus tires, would represent a logical next step. It is concluded that the objective was achieved: it has been demonstrated that it is possible to create high-performance predictive models for tire sales in Brazil using public data and consolidated statistical techniques, generating applicable knowledge for strategic decision-making in the industrial sector.
References:
Automotive Manufacturers Association [ANFAVEA]. 2025. Automobile sales reports. Available at: https://anfavea. com. br/site/edicoes-em-excel/ . Accessed on: 01 Jul. 2025.
National Tire Industry Association [ANIP]. 2025. Tire sales reports. Available at: https://www. anip. org. br/anip-em-numeros/publicacoes/ . Accessed on: 01 Jul. 2025.
Comex Stat. 2025. Import statistics for tires, natural rubber, and tire cord. Available at: https://www. gov. br/mdic/pt-br/assuntos/comercio-exterior/estatisticas/base-de-dados-bruta. Accessed on: 01 Jul. 2025.
National Confederation of Industry [CNI]. 2025. Indicators. Available at: https://indicadores. sistemaindustria. com. br/indicadores/externo/estatisticaAcessoSistemaExterno. faces . Accessed on: 01 Jul. 2025.
Fávero, Luis Paulo; Belfiore, Patrícia. 2024. Manual de Análise de Dados: estatística e machine learning com Excel®, SPSS®, R® e Python®. 2nd ed. LTC, Rio de Janeiro, Brazil.
Grus, Joel. 2019. Data Science from Scratch. 2nd ed. O’Reilly, Sebastopol, USA.
Hyndman, R. J., Athanasopoulos, G. 2021. Forecasting: principles and practice. 3rd edition. OTexts, Melbourn, Australia. Available at: < https://otexts. com/fpppg/arima-estima%C3%A7%C3%A3o. html >. Accessed on: 01 Jul 2025.
Joseph, Manu; Tackes, Jeffrey. 2024. Modern Time Series Forecasting with Python. 2nd ed. Packt, Birmingham, United Kingdom.
Morettin, Pedro A. 2018. Análise de Séries Temporais. 1st ed. Editora Blusher, São Paulo, Brazil.
Nielsen, Aileen. 2021. Análise Prática de Séries Temporais: predição com estatística e aprendizado de máquina. 1st ed. Alta Books, Rio de Janeiro, Brazil.
Oliveira et al. 2024. Introdução às séries temporais: uma abordagem prática em Python. 1st ed. Editora Mackenzie, São Paulo, Brazil.
Serasa. 2025. Default indicators. Available at: <https://www. serasaexperian. com. br/conteudos/indicadores-economicos/ >. Accessed on: 05 Jul 2025.
Taylor, Brent S. 2015. Utilizing ANNs to Improve the Forecast for Tire Demand. In partial fulfillment of the requirements for the degree Master of Science. Russ College of Engineering and Technology of Ohio University, Ohio University, Ohio, USA.
VanderPlas, Jake. 2016. Python Data Science Handbook. Essential Tools for Working with Data. 1st ed. O’Reilly, Sebastopol, USA.
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: