A carregar...
Publicação
Zero-Shot Learning for Multivariate Time Series Forecasting: Advancing Sales Predictions with Deep Neural Networks
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 1.93 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Time series forecasting is a critical tool for supporting strategic decisions across sectors
such as agriculture, finance, and energy. This dissertation aims to compare the
performance of recently developed forecasting models with well-established ones,
applied to a multivariate time series of sales data for an agrochemical product. The
analysis includes exogenous variables of climatic and economic nature, assessing the
models’ ability to integrate external factors into prediction improvement. Three
baseline models (LSTM, GRU, and N-HiTS) and two zero-shot capable models (TimeGPT
and Tiny Time Mixers) were tested in both their default and fine-tuned configurations.
Model performance was evaluated using metrics such as MAE, RMSE, SMAPE, and
MASE. Results show that zero-shot models, particularly TimeGPT, deliver competitive
performance with significant advantages in speed, usability, and the absence of
hyperparameter tuning. When used with exogenous variables, fine-tuned models
showed noticeable accuracy improvements, although the computational cost was not
always justified. The findings suggest that for high-volume, low-margin products, zeroshot models may offer the most efficient solution. The study also highlights the
importance of model interpretability and proposes future research directions focused
on integrating explainability techniques to better understand the drivers behind the
forecasts. This dissertation contributes practical insights into the use of advanced
forecasting models and raises important considerations about the trade-off between
performance, complexity, and applicability in real business contexts.
Descrição
Dissertation presented as the partial requirement for obtaining a Master's degree in Information Management, specialization in Business Intelligence
Palavras-chave
Time Series Forecasting Deep Learning Zero-Shot forecasting Multivariate Time-Series Sales Forecasting SDG 8 - Decent work and economic growth SDG 9 - Industry, innovation and infrastructure SDG 12 - Responsible production and consumption