A carregar...
Publicação
Modelos de previsão baseados em séries temporais: uma abordagem comparativa aplicada à gestão operacional
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 4.2 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Este projeto, desenvolvido no âmbito do Mestrado em Ciência de Dados para Empresas, teve como objetivo principal a realização de uma análise comparativa entre diferentes modelos de previsão de séries temporais, aplicados a dados históricos do sistema ATX (2021–2024). Foram avaliados cinco modelos: Seasonal Autoregressive Integrated Moving Average(SARIMA), Holt-Winters (método de suavização exponencial), Neural Basis Expansion Analysis for Time Series (N-BEATS), Deep Autoregressive Model (DeepAR) e Prophet. A metodologia adotada seguiu o modelo Cross Industry Standard Process for Data Mining(CRISP-DM), abrangendo desde a compreensão do negócio até à avaliação e implementação dos modelos. O processo incluiu a preparação rigorosa dos dados, a aplicação de técnicas estatísticas e de aprendizagem profunda, e a validação através de métricas de erro como Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE) e Mean Absolute Percentage Error (MAPE), complementadas pelo teste estatístico de Diebold-Mariano. Os resultados indicaram que o modelo Prophet apresentou o melhor desempenho global, destacando-se pela sua robustez, consistência e adequação em contextos empresariais. Este estudo evidencia a importância da previsão precisa para a eficiência operacional e o planeamento de recursos, propondo a integração dos modelos mais eficazes em sistemas internos, com recurso a dashboards interativos e Application Programming Interface (API). Recomenda-se, ainda, a continuação da investigação com modelos híbridos e a inclusão de variáveis exógenas, de forma a melhorar a precisão das previsões em cenários futuros.
This project, developed as part of the Master's in Data Science for Business, aimed to conduct a comparative analysis of different time series forecasting models applied to historical data from the ATX system (2021–2024). Five models were evaluated: SARIMA (Seasonal Autoregressive Integrated Moving Average), Holt-Winters (exponential smoothing method), N BEATS (Neural Basis Expansion Analysis for Time Series), DeepAR (Deep Autoregressive Model), and Prophet. The methodology adopted followed the Cross Industry Standard Process for Data Mining(CRISP-DM) framework, encompassing everything from business understanding to model evaluation and implementation. The process included rigorous data preparation, the application of statistical and deep learning techniques, and validation using error metrics such as MAE (Mean Absolute Error), MSE (Mean Squared Error), RMSE (Root Mean Squared Error), and MAPE (Mean Absolute Percentage Error), complemented by the Diebold-Mariano statistical test. The results indicated that the Prophet model performed best overall, standing out for its robustness, consistency, and suitability in business contexts. This study highlights the importance of accurate forecasting for operational efficiency and resource planning, proposing the integration of the most effective models into internal systems, using interactive dashboards and APIs. We also recommend further research with hybrid models and the inclusion of exogenous variables to improve forecast accuracy in future scenarios.
This project, developed as part of the Master's in Data Science for Business, aimed to conduct a comparative analysis of different time series forecasting models applied to historical data from the ATX system (2021–2024). Five models were evaluated: SARIMA (Seasonal Autoregressive Integrated Moving Average), Holt-Winters (exponential smoothing method), N BEATS (Neural Basis Expansion Analysis for Time Series), DeepAR (Deep Autoregressive Model), and Prophet. The methodology adopted followed the Cross Industry Standard Process for Data Mining(CRISP-DM) framework, encompassing everything from business understanding to model evaluation and implementation. The process included rigorous data preparation, the application of statistical and deep learning techniques, and validation using error metrics such as MAE (Mean Absolute Error), MSE (Mean Squared Error), RMSE (Root Mean Squared Error), and MAPE (Mean Absolute Percentage Error), complemented by the Diebold-Mariano statistical test. The results indicated that the Prophet model performed best overall, standing out for its robustness, consistency, and suitability in business contexts. This study highlights the importance of accurate forecasting for operational efficiency and resource planning, proposing the integration of the most effective models into internal systems, using interactive dashboards and APIs. We also recommend further research with hybrid models and the inclusion of exogenous variables to improve forecast accuracy in future scenarios.