| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 5.86 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O sector do retalho, avaliado em biliões de euros, é dinâmico e em constante evolução, enfrentando
desafios na previsão precisa da procura devido à rápida mudança nas preferências dos
consumidores, às tendências emergentes e à variedade de atributos dos produtos. A necessidade
de planear as vendas com antecedência e a escassez de dados concretos aumentam ainda mais a
complexidade. Métodos de previsão como os estatísticos, apreciados pela sua clareza, podem ser
insuficientes com dados complexos, enquanto as técnicas de aprendizagem automática conseguem
identificar padrões mais complexos, apesar de exigirem grandes volumes de dados.
Esta dissertação utiliza a metodologia CRISP-DM para explorar métodos estatísticos de previsão e
a adaptabilidade dos algoritmos de aprendizagem automática na previsão de vendas no setor de
retalho. A combinação destas técnicas avançadas com o processo CRISP-DM proporciona uma
análise mais robusta e flexível, aumentando a eficiência e a precisão das previsões de procura. Isto
permite otimizar a gestão de stocks e o planeamento estratégico, o que pode melhorar
significativamente o desempenho operacional e a rentabilidade da empresa.
Os dados analisados pertencem à empresa norte-americana Walmart e abrangem as vendas de 10
lojas distribuídas por três estados: Califórnia, Texas e Wisconsin, ao longo de um período de 5 anos.
Para prever as vendas futuras, foram desenvolvidos diversos modelos, incluindo um modelo global,
treinado com todos os dados, e outros específicos por loja e por estado. O modelo global captura
tendências gerais, enquanto os modelos específicos por estado e por loja identificam padrões
únicos de cada região e unidade, proporcionando previsões mais ajustadas a nível local.
Adicionalmente, cada modelo de aprendizagem automática foi treinado em 4 cenários distintos:
utilizando dados totais, do último ano, dos últimos 3 meses e do último mês, com o objetivo de
observar o impacto das variações temporais nas previsões. Foram também considerados fatores
como feriados e dias da semana nos modelos, que têm um impacto significativo nas vendas.
Os resultados obtidos mostram que, entre os métodos estatísticos, o modelo agregado por loja com
SARIMA foi o mais eficaz para a previsão de vendas em contextos sazonais, alcançando um MAE de
224 e um MAPE de 5,39%. No que diz respeito aos métodos de aprendizagem automática, o Support
Vector Machines (SVM) destacou-se ao apresentar valores elevados de R², variando entre 77% e
80% para os modelos agregados por estado e loja, com erros relativamente baixos. A análise
também revelou que o modelo agregado global apresentou um desempenho inferior, refletindo
uma capacidade limitada para capturar as particularidades regionais e locais. Estes resultados
sublinham a importância de utilizar dados atualizados e ajustar os modelos às recentes variações
do mercado, enfatizando a necessidade de uma abordagem adaptativa na modelagem preditiva. A
comparação entre diferentes metodologias e a inclusão de hiperparâmetros foram cruciais para a
precisão das previsões, com os resultados variando conforme o modelo e o contexto de aplicação.
The retail sector, valued in billions of euros, is dynamic and constantly evolving, facing challenges in accurately forecasting demand due to rapid changes in consumer preferences, emerging trends, and the variety of product attributes. The need to plan sales in advance and the lack of concrete data further increase the complexity. Forecasting methods like statistical ones, valued for their clarity, may be insufficient with complex data, while machine learning techniques can identify more complex patterns, although they require large volumes of data. This dissertation uses the CRISP-DM methodology to explore statistical forecasting methods and the adaptability of machine learning algorithms in sales forecasting within the retail sector. Combining these advanced techniques with the CRISP-DM process provides a more robust and flexible analysis, enhancing the efficiency and accuracy of demand forecasts. This enables optimization of inventory management and strategic planning, potentially significantly improving operational performance and company profitability. The analyzed data comes from the American company Walmart and covers sales from 10 stores across three states: California, Texas, and Wisconsin, over a period of 5 years. Various models were developed to forecast future sales, including a global model trained with all data, and other models specific to each store and state. The global model captures general trends, while the state-specific and store-specific models identify unique patterns of each region and unit, providing more localized forecasts. Additionally, each machine learning model was trained under 4 different scenarios: using total data, data from the last year, the last 3 months, and the last month, aiming to observe the impact of temporal variations on forecasts. Factors such as holidays and days of the week were also considered in the models, as they have a significant impact on sales. The results showed that among statistical methods, the store-aggregated model with SARIMA was the most effective for sales forecasting in seasonal contexts, achieving an MAE of 224 and a MAPE of 5.39%. Regarding machine learning methods, Support Vector Machines (SVM) stood out by presenting high R² values, ranging from 77% to 80% for state and store-aggregated models, with relatively low errors. The analysis also revealed that the global aggregated model performed poorly, reflecting a limited ability to capture regional and local specifics. These results highlight the importance of using updated data and adjusting models to recent market variations, emphasizing the need for an adaptive approach in predictive modeling. Comparing different methodologies and including hyperparameters were crucial for forecast accuracy, with results varying depending on the model and application context.
The retail sector, valued in billions of euros, is dynamic and constantly evolving, facing challenges in accurately forecasting demand due to rapid changes in consumer preferences, emerging trends, and the variety of product attributes. The need to plan sales in advance and the lack of concrete data further increase the complexity. Forecasting methods like statistical ones, valued for their clarity, may be insufficient with complex data, while machine learning techniques can identify more complex patterns, although they require large volumes of data. This dissertation uses the CRISP-DM methodology to explore statistical forecasting methods and the adaptability of machine learning algorithms in sales forecasting within the retail sector. Combining these advanced techniques with the CRISP-DM process provides a more robust and flexible analysis, enhancing the efficiency and accuracy of demand forecasts. This enables optimization of inventory management and strategic planning, potentially significantly improving operational performance and company profitability. The analyzed data comes from the American company Walmart and covers sales from 10 stores across three states: California, Texas, and Wisconsin, over a period of 5 years. Various models were developed to forecast future sales, including a global model trained with all data, and other models specific to each store and state. The global model captures general trends, while the state-specific and store-specific models identify unique patterns of each region and unit, providing more localized forecasts. Additionally, each machine learning model was trained under 4 different scenarios: using total data, data from the last year, the last 3 months, and the last month, aiming to observe the impact of temporal variations on forecasts. Factors such as holidays and days of the week were also considered in the models, as they have a significant impact on sales. The results showed that among statistical methods, the store-aggregated model with SARIMA was the most effective for sales forecasting in seasonal contexts, achieving an MAE of 224 and a MAPE of 5.39%. Regarding machine learning methods, Support Vector Machines (SVM) stood out by presenting high R² values, ranging from 77% to 80% for state and store-aggregated models, with relatively low errors. The analysis also revealed that the global aggregated model performed poorly, reflecting a limited ability to capture regional and local specifics. These results highlight the importance of using updated data and adjusting models to recent market variations, emphasizing the need for an adaptive approach in predictive modeling. Comparing different methodologies and including hyperparameters were crucial for forecast accuracy, with results varying depending on the model and application context.
Description
Keywords
CRISP-DM Forecasting Statistical methods Machine learning Previsão Métodos estatísticos