| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 5.86 MB | Adobe PDF |
Advisor(s)
Abstract(s)
O presente trabalho aborda a criação de um modelo de previsão de vendas semanais por
artigo / localização com base em dados reais de uma empresa de retalho líder no setor de
cosmética.
O modelo atualmente utilizado pela empresa apresenta indícios de não captar corretamente
as características da procura dos seus clientes, o que gera ineficiências dentro da cadeia de
abastecimento por excesso de alocação de stock a um determinado ponto, assim como falhas
de stock e consequentes perdas de venda em alturas sazonais ou em que existem eventos
especiais.
Foram aplicados métodos clássicos de previsão com base em séries temporais, como o
método de Amortecimento Exponencial Simples, o Método de Holt, Holt-Winters, e o método
ARIMA. Adicionalmente, foi também aplicado um algoritmo de multilayer perceptron (MLP) que
utiliza variáveis de eventos culturais, nacionais, desportivos, e meteorológicos, além de
informação sobre a família e subfamília do artigo. Para estimar os valores da sazonalidade, da
tendência e os resíduos de cada série temporal, é utilizado o Amortecimento Exponencial
Simples para prever os pontos do horizonte de previsão e aplicado o método STL com
sazonalidade a um ano para extrair os valores estimados, que são posteriormente aplicados na
previsão de vendas gerada pelo modelo MLP.
Os modelos foram testados utilizando um método de Janela Deslizante ao longo de um ano,
em que a cada ponto se procurava prever os dois pontos futuros, utilizando o RMSE médio de
cada modelo como avaliação. O modelo MLP conseguiu o melhor desempenho em 16 dos 19
artigos utilizados para teste, o que demonstra a potencialidade da utilização de algoritmos de
Machine Learning para prever a procura de uma cadeia de retalho.
This study addresses the development of a sales forecasting model for weekly sales by item/location based on real data from a leading retail company in the cosmetics sector. The model currently used by the company shows signs of not accurately capturing the demand characteristics within the supply chain, leading to inefficiencies due to overstocking at specific points, as well as stock shortages and subsequent sales losses during seasonal peaks or special events. Classic forecasting methods based on time series were applied, such as the Simple Exponential Smoothing method, the Holt Method, Holt-Winters, and the ARIMA method. Additionally, a multilayer perceptron (MLP) algorithm was employed, incorporating variables from cultural, national, sports, and weather events, along with information about the item's family and subfamily. To estimate the seasonality, trend, and residuals of each time series, the Simple Exponential Smoothing method is used to predict the forecast horizon points, and the STL method with a one-year seasonality is applied to extract the estimated values, which are then utilized in the sales forecast generated by the MLP model. The models were tested using a Rolling Window method over a year, where each point aimed to predict the next two future points, using the average RMSE of each model as an evaluation metric. The MLP model achieved the best performance on 16 out of the 19 items tested, demonstrating the potential of using Machine Learning algorithms to forecast the demand in a retail chain.
This study addresses the development of a sales forecasting model for weekly sales by item/location based on real data from a leading retail company in the cosmetics sector. The model currently used by the company shows signs of not accurately capturing the demand characteristics within the supply chain, leading to inefficiencies due to overstocking at specific points, as well as stock shortages and subsequent sales losses during seasonal peaks or special events. Classic forecasting methods based on time series were applied, such as the Simple Exponential Smoothing method, the Holt Method, Holt-Winters, and the ARIMA method. Additionally, a multilayer perceptron (MLP) algorithm was employed, incorporating variables from cultural, national, sports, and weather events, along with information about the item's family and subfamily. To estimate the seasonality, trend, and residuals of each time series, the Simple Exponential Smoothing method is used to predict the forecast horizon points, and the STL method with a one-year seasonality is applied to extract the estimated values, which are then utilized in the sales forecast generated by the MLP model. The models were tested using a Rolling Window method over a year, where each point aimed to predict the next two future points, using the average RMSE of each model as an evaluation metric. The MLP model achieved the best performance on 16 out of the 19 items tested, demonstrating the potential of using Machine Learning algorithms to forecast the demand in a retail chain.
Description
Keywords
Forecasting Retail Supply Chain Management Stock Machine Learning