An Automated Meta- Forecasting Framework For Intelligent Model Selection in Retail Time Series Analysis

Abstract

The inability to predict the suitability of a forecasting model for each and every sales series makes retail sales forecasting difficult. This research introduces an intelligent meta-forecasting framework that integrates data-driven model recommendation, quantitative forecasting, retrieval-augmented generation (RAG) and large language model (LLM) decision support. Experiments were carried out with 10,194 retail transactions made between 2019-2022. The results showed that there are 17 sub-category sales series, which led to the creation of 68 forecasting tasks for the evaluation of the meta-learning. When tested on unseen retail series, the Random Forest meta-classifier outperformed the other models with a model-selection accuracy of 40.0% and a macro-F1 score of 0.273. The mean test RMSE was 259.349, which was significantly better than random selection (RMSE 316.931; regret 68.997), but was not as good as a global baseline of SGD (RMSE 250.967; regret 3.032). A prototype RAG–LLM system is also instantiated with semantic retrieval, model recommendation and interpretation of forecasts through the conversation. The results show the suitability of a forecasting model to be heterogeneous and also highlight the limitation of cross-series generalization in automated meta-learning.