Análise preditiva de emplacamentos de implementos rodoviários usando modelos de séries temporais e redes neurais : estudo de caso comdados da ANFIR(2013-2024)

Abstract

This study proposed the development of a predictive model for the registration of road implements in Brazil, based on historical data provided by ANFIR for the period from 2013 to 2024. The research adopted a hybrid approach, combining traditional statistical time series techniques such as ARIMA and SARIMA with contemporary machine learning and artificial intelligence methods, including Artificial Neural Networks (ANN), Deep Learning, Naive Bayes, Random Forest, Multiple Linear Regression (MLR), Simple Linear Regression (SLR), Exponential Smoothing, and XGBoost. The primary objective was to identify seasonal patterns, historical trends, and nonlinear relationships among multiple variables that influence the behavior of the implement market.The analysis was conducted using historical registration data for Trailers/Semi-trailers and Truck Bodies, with an initial assessment covering the full period from 2013 to 2023. This segmentation enabled a comprehensive understanding of registration trends, taking into account key economic events such as the 2014-2016 recession, marked by GDP contractions of -3.5% and -3.3% (according to IBGE), and the COVID-19 pandemic in 2020, which severely affected production and sales in the sector. Due to these disruptive events, the data were further segmented into two additional timeframes: from 2018 to 2023, to capture more recent trends, and from 2021 to 2023, to isolate the effects of the pandemic.The forecasts generated by each model were compared to the actual 2024 data using evaluation metrics such as Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE). In the full period analysis (2013-2023), the SLR model stood out as the most effective for predicting semi-trailer registrations, achieving strong performance metrics (MAPE = 5.93%; MAE = 437.33; RMSE = 494.60), proving particularly effective in contexts with regular seasonality. Exponential Smoothing also performed well (MAPE = 6.57%) by leveraging the linear trend in the historical data. In the post-pandemic period (2021-2023), SLR again emerged as the best-performing model (MAPE = 5.21%; MAE = 370.30; RMSE = 472.55), reflecting the predominance of a linear trend during this shorter interval. The Deep Learning model also delivered satisfactory results (MAPE = 5.36%), though its higher computational cost limits practical applicability. Based on these results, it is recommended to use the SLR model for long-term forecastsof semitrailer registrations, complemented by Exponential Smoothing to capture specific seasonal effects. For more recent periods, SLR is suggested when a clear linear trend is present; otherwise, Deep Learning demonstrates greater effectiveness. In the case of truck bodies, models such as ARIMA (2013-2023, MAPE = 14.96%), MLR (2018-2023, MAPE = 9.37%), and Deep Learning (2021-2023, MAPE = 13.75%) proved to be better suited to the variability in the data for those specific timeframes. Finally, the importance of continuous monitoring and periodic recalibration of the models is emphasized, as unexpected events-such as economic crises or pandemics-can significantly alter historical patterns and compromise predictive accuracy. [resumo fornecido pelo autor]