Banca de DEFESA: ESDRAS ALEX FREIRE DE OLIVEIRA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : ESDRAS ALEX FREIRE DE OLIVEIRA
DATE: 27/09/2024
TIME: 10:00
LOCAL: Plataforma Google Meet - Link https://meet.google.com/aor-waop-ftx
TITLE:

Forecasting time series of products sold at retail

KEY WORDS:

Forecasting. Time series. Stacked generalization. Machine learning. Product families.

PAGES: 140
BIG AREA: Ciências Agrárias
AREA: Agronomia
SUMMARY:

Forecasting sales demand is essential for strategic planning and efficient supply chain management. This practice allows organizations to anticipate market needs, forming the basis for financial planning, inventory control, and product acquisition, thus reducing costs. In this context, the adoption of Machine Learning (ML) models for time series forecasting emerges as a valuable tool. In this research, the objective was to analyze the predictive capacity of individual and combined mathematical models through a stacked generalization approach, for forecasting time series of product sales volume in the retail segment. The individual models were trained and, subsequently, the five best forecasts among the ETS, SARIMA, k-NN, SVR, DTR, RFR, ETR, XGBoost, LightGBM, MLP, and LSTM models were selected based on their performance in the MAAPE measure. The combined models used the outputs of the five base models selected in the first stage. Next, the four best combiners for each time series were chosen, including those based on the simple mean (cSA), the simple median (cSM) and the minimum variance (cMV). The experiments involving the crossover between models resulted in 165 combinations of base models and 132 combinations of metamodels for the set of time series analyzed in this research. Then, a performance comparison of the proposed models was performed using MAAPE, MAE and ME. The comparative analysis of the demand time series showed that the combined models outperformed the individual models, with MAAPE ranging from 6% to 46%. In comparison with other studies, the method used in this research presented superior performance, with MAE ranging from 28.4 to 70 units in some product categories, while in others the performance was inferior, with MAE between 128.9 and 357.4 units. It is worth noting that the cSVR and cLSTM stacking models, at the store level, outperformed other models, such as SARIMA, ANN, CNN and the combined CNN+LSTM. While the MAPE of these models ranged from 12.9% to 17.6%, those of cSVR and cLSTM stacking were below 11.6%, highlighting the effectiveness of this approach in forecasting time series with high variability.

COMMITTEE MEMBERS:
Externo à Instituição - BERNARDO SOBRINHO SIMÕES ALMADA-LOBO
Interno - ***.218.133-** - CÍCERO CARLOS FELIX DE OLIVEIRA - IFCE
Interno - FRANCISCO ALIXANDRE AVILA RODRIGUES
Presidente - PAULO RENATO ALVES FIRMINO