Loading...
Publication
Transfer learning approaches for Business Forecasting
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 3.27 MB | Adobe PDF |
Authors
Abstract(s)
Esta dissertação examina a forma como a aprendizagem por transferência pode melhorar a previsão do negócio de retalho, melhorando a precisão e a eficiência dos modelos de previsão da procura, projeção de vendas e otimização de inventário. Aborda dois desafios principais na previsão de retalho: escassez de dados e generalização. A aprendizagem por transferência utiliza modelos pré-treinados de áreas relacionadas para construir modelos de previsão mais robustos, integrando conhecimentos de diversos conjuntos de dados para criar soluções adaptáveis a vários ambientes de retalho. O estudo compara cinco modelos de aprendizagem por transferência: Chronos da AWS, TimeGPT da Nixtla, Lag-Llama da Invenia Technical Computing, TimesFM da Google Research e Uni2TS da SalesforceAI, avaliados utilizando o conjunto de dados M5, que integra dados reais de vendas a retalho da Walmart. Este conjunto de dados apresenta desafios devido às suas estruturas hierárquicas, dependências temporais e factores externos como promoções e flutuações de preços. O desempenho de cada modelo é avaliado utilizando o erro médio absoluto escalado (MASE) e a pontuação de probabilidade classificada contínua (CRPS).
O MASE mede a exatidão das previsões pontuais, permitindo comparações entre séries cronológicas com sazonalidade variável, enquanto o CRPS avalia a exatidão das previsões probabilísticas, crucial para compreender a incerteza na procura no retalho. Os modelos são testados em dois cenários: Zero-Shot, em que os modelos são aplicados sem treino no conjunto de dados M5, e Fine-Tuning, em que os modelos são especificamente ajustados ao mesmo. Os resultados mostram que o TimesFM supera consistentemente os outros modelos em ambos os cenários, demonstrando robustez e versatilidade. O Lag-Llama tem um bom desempenho nas definições Zero-Shot, captando eficazmente os padrões gerais das séries temporais. Os modelos Chronos melhoram com o tamanho, mas permanecem menos competitivos, enquanto o Uni2TS enfrenta desafios no Fine-Tuning, indicando um potencial sobreajuste.
This dissertation examines how transfer learning can enhance retail business forecasting, improving the accuracy and efficiency of models for demand forecasting, sales projection, and inventory optimization. It addresses two main challenges in retail forecasting: data scarcity and generalization. Transfer learning leverages pre-trained models from related fields to build more robust forecasting models, integrating insights from diverse datasets to create adaptable solutions for various retail environments. The study compares five transfer learning models: Chronos by AWS, TimeGPT by Nixtla, Lag-Llama by Invenia Technical Computing, TimesFM by Google Research, and Uni2TS by SalesforceAI, evaluated using the M5 dataset, which contains real-world retail sales data from Walmart. This dataset poses challenges due to its hierarchical structures, temporal dependencies, and external factors like promotions and price fluctuations. Model performance is assessed using Mean Absolute Scaled Error (MASE) and Continuous Ranked Probability Score (CRPS). MASE measures point forecast accuracy, allowing comparisons across time series with varying seasonality, while CRPS evaluates probabilistic forecast accuracy, crucial for understanding uncertainty in retail demand. The models are tested in two settings: Zero-Shot, where they are applied without training on the M5 dataset, and Fine-Tuning, where they are specifically adjusted to it. Results show that TimesFM consistently outperforms the other models in both settings, demonstrating robustness and versatility. Lag-Llama performs well in Zero-Shot settings, capturing general time series patterns effectively. Chronos models improve with size but remain less competitive, while Uni2TS faces challenges in Fine-Tuning, indicating potential overfitting.
This dissertation examines how transfer learning can enhance retail business forecasting, improving the accuracy and efficiency of models for demand forecasting, sales projection, and inventory optimization. It addresses two main challenges in retail forecasting: data scarcity and generalization. Transfer learning leverages pre-trained models from related fields to build more robust forecasting models, integrating insights from diverse datasets to create adaptable solutions for various retail environments. The study compares five transfer learning models: Chronos by AWS, TimeGPT by Nixtla, Lag-Llama by Invenia Technical Computing, TimesFM by Google Research, and Uni2TS by SalesforceAI, evaluated using the M5 dataset, which contains real-world retail sales data from Walmart. This dataset poses challenges due to its hierarchical structures, temporal dependencies, and external factors like promotions and price fluctuations. Model performance is assessed using Mean Absolute Scaled Error (MASE) and Continuous Ranked Probability Score (CRPS). MASE measures point forecast accuracy, allowing comparisons across time series with varying seasonality, while CRPS evaluates probabilistic forecast accuracy, crucial for understanding uncertainty in retail demand. The models are tested in two settings: Zero-Shot, where they are applied without training on the M5 dataset, and Fine-Tuning, where they are specifically adjusted to it. Results show that TimesFM consistently outperforms the other models in both settings, demonstrating robustness and versatility. Lag-Llama performs well in Zero-Shot settings, capturing general time series patterns effectively. Chronos models improve with size but remain less competitive, while Uni2TS faces challenges in Fine-Tuning, indicating potential overfitting.
Description
Keywords
Transfer learning Business forecasting Natural language processing Transformer architecture Deep neural networks Attention-based models Previsão no negócio Processamento de linguagem natural Arquitetura transformer Redes neurais Modelos baseados no mecanismo de atenção