Forecasting time series with negative R²: an application in estimating spare parts demand ; Previsão de séries temporais com R² negativo: uma aplicação na estimativa da procura de peças sobresselentes

Forecasting sales of replacement products presents a number of difficulties due to the unavailability of demand. The main obstacles are the intermittency of sales and the quantities required. The objective of this project is to conduct sales forecasting using machine learning models, with a focus on the distinction between a conventional forecast and a split forecast. The products under analysis exhibited a prolonged period of no sales, which presented a challenge for machine learning models in identifying patterns. Additionally, the time series exhibited a high number of zeros, indicating a scarcity of demand. A model was created that makes different predictions between weeks with sales and weeks without sales, and is then compared with classic machine learning models. The results demonstrated that the new model exhibited notable enhancements in forecasting accuracy, particularly in forecasting weeks without sales. This was evident in the traditional model, which tended to make forecasts closer to the average, assuming a consistent demand. In order to evaluate the forecasts, the RMSE and MAE metrics were employed extensively in order to emphasise the number of weeks without sales, as this was the core problem. ; A previsão de vendas de produtos de substituição apresenta várias dificuldades devido à impressibilidade da procura, os principais obstáculos passam pela intermitência das vendas assim como pelas quantidades requisitadas. O objetivo deste projeto passa por realizar a previsão das vendas por meio de modelos de Machine Learning, estudando a diferença entre uma previsão clássica e de uma previsão repartida. Os produtos em análise continham enumeras semanas sem vendas, dificultando a identificação de padrões por parte dos modelos de Machine Learning, as séries temporais eram caraterizadas pelo elevado número de zeros, demonstrando a escassez de procura. Foi criado um modelo que realiza previsões distintas, entre as semanas com vendas e as semanas sem vendas, sendo posteriormente comparado com os modelos de ...