Loading...
Publication
Modelação preditiva baseada em algoritmos de ia para previsão de vendas no retalho
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 6.87 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O lançamento de artigos sem histórico de vendas aumenta o risco operacional no retalho. Esta
investigação propõe uma estrutura preditiva que estima a procura inicial de novos produtos a partir
de 3 278 907 transações reais (2022-2024) cobrindo 14 422 SKUs, 120 lojas e dois segmentos de
cliente. Seguindo o ciclo CRISP-DM, procedeu-se à limpeza dos dados, análise exploratória e
engenharia de variáveis, antes de comparar quatro algoritmos — XGBoost, LightGBM, LSTM e
Transformer — em cenários global e por família de produtos, avaliados com MAE, RMSE, MAPE e
R².
Os resultados revelam dois patamares distintos: os modelos de árvores de gradiente (XGBoost ≈
LightGBM) registam erros médios substancialmente menores e R² positivos, ao passo que as redes
neuronais sequenciais (LSTM, Transformer) apresentam elevada variabilidade e R² negativos em
várias famílias. O XGBoost treinado globalmente demonstra o menor RMSE ponderado e o melhor
equilíbrio viés-variância, sendo recomendado como motor de previsão único para toda a gama de
artigos. Esta solução simplifica a operação, mantém precisão elevada e foi integrada num protótipo
de dashboard web para validação em contexto real.
Conclui-se que um modelo único, alimentado pela diversidade de SKUs e lojas, generaliza padrões
de procura com eficácia, oferecendo uma ferramenta prática para apoiar decisões ao nível do
portefólio e de planeamento comercial e operacional.
Introducing retail items with no sales history entails significant demand uncertainty. This study develops a predictive framework that leverages 3,278,907 real transactions (2022–2024) spanning 14,422 SKUs, 120 stores and two customer segments. Adhering to the CRISP-DM methodology, the workflow comprises data cleansing, exploratory analysis, feature engineering and the assessment of four algorithms — XGBoost, LightGBM, LSTM and Transformer— under global and family-specific settings using MAE, RMSE, MAPE and R². Findings disclose two performance tiers: gradient-boosted trees (XGBoost ≈ LightGBM) achieve markedly lower errors and positive R², whereas sequence models (LSTM, Transformer) struggle with variance and often yield negative R². The globally trained XGBoost secures the lowest weighted RMSE and the best bias-variance trade-off and is thus recommended as a single forecasting engine for the entire product range. A web-based dashboard prototype demonstrates real-time deployment and business applicability. The study concludes that a unified model, trained across heterogeneous SKUs and outlets, can generalize demand patterns effectively, delivering high accuracy with reduced operational complexity for assortment and inventory planning.
Introducing retail items with no sales history entails significant demand uncertainty. This study develops a predictive framework that leverages 3,278,907 real transactions (2022–2024) spanning 14,422 SKUs, 120 stores and two customer segments. Adhering to the CRISP-DM methodology, the workflow comprises data cleansing, exploratory analysis, feature engineering and the assessment of four algorithms — XGBoost, LightGBM, LSTM and Transformer— under global and family-specific settings using MAE, RMSE, MAPE and R². Findings disclose two performance tiers: gradient-boosted trees (XGBoost ≈ LightGBM) achieve markedly lower errors and positive R², whereas sequence models (LSTM, Transformer) struggle with variance and often yield negative R². The globally trained XGBoost secures the lowest weighted RMSE and the best bias-variance trade-off and is thus recommended as a single forecasting engine for the entire product range. A web-based dashboard prototype demonstrates real-time deployment and business applicability. The study concludes that a unified model, trained across heterogeneous SKUs and outlets, can generalize demand patterns effectively, delivering high accuracy with reduced operational complexity for assortment and inventory planning.
Description
Keywords
sales forecasting new products XGBoost retail machine learning artificial intelligence Inteligência artificial Novos produtos no retalho Previsão de vendas