A carregar...
Publicação
Comparative Analysis of Classical Models and Foundation Models for Retail Sales Forecasting
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.8 MB | Adobe PDF |
Orientador(es)
Resumo(s)
This study evaluates and compares the performance of classical forecasting models and
advanced machine learning approaches for retail sales forecasting. Specifically, the analysis
focuses on the applicability of foundation models, namely TimeGPT-1 and Moirai, in addition
to Seasonal Autoregressive Integrated Moving Average (SARIMA), Holt-Winters, and Prophet,
across three distinct retail product categories, namely camcorders, media tablets, and toys.
The models were applied to weekly sales data over two distinct forecasting windows: MarchApril 2024, representing stable demand, and November-December 2024, characterised by
heightened volatility due to the influence of holiday and promotional activities. Forecast
accuracy was assessed using Root Mean Squared Error (RMSE) and Mean Absolute Percentage
Error (MAPE). The empirical results demonstrate that foundation models outperformed
classical models in terms of absolute accuracy, particularly in contexts characterised by
volatile demand conditions. TimeGPT-1 demonstrated a consistent capacity to generate
reliable and consistent forecasts across categories. Moirai demonstrated notable efficacy in
moderate volatility environments. Prophet produced stable baseline forecasts, but the
addition of external regressors did not systematically improve performance. The classical
models maintained their competitiveness in stable seasonal contexts. These findings
underscore the limitations of classical forecasting models in dynamic retail environments and
highlight the potential of foundation models to generalise across different products and
timeframes.
Descrição
Dissertation presented as the partial requirement for obtaining a Master's degree in Statistics and Information Management, specialization in Information Analysis and Management
Palavras-chave
Forecasting Foundation Models Methodological Comparison Model Performance Univariate Time Series Holiday and Promotion Impact