LARGE LANGUAGE MODELS FOR TIME SERIES ANALYSIS IN RETAIL: PERSPECTIVES AND OPPORTUNITIES

Abstract

The article presents a comprehensive overview of state-of-the-art methodologies in the retail industry, framed within the context of Retail 4.0. It delves into the usage of various cutting-edge technologies, including large language models (LLMs), machine learning algorithms, statistical methods, and econometrics models. Specifically, the study explores the potential of LLMs in conducting time series analysis, identifying three primary avenues of their application: as supportive tools, forecasting mechanisms utilizing zero-shot learning principles, and next-generation agents with interactive functionality. Each application direction is examined, detailing their respective advantages and drawbacks, alongside highlighting the existing gaps or constraints in empirical research within the field. The research includes empirical experiments conducted with the TimeGPT model, a foundational model tailored for time series analysis. This model demonstrated competitiveness when compared against classical methodologies. However, the study emphasizes the necessity of expanding experimentation to collect more information to evaluate the solution in the context of sales forecasting tasks. Moreover, the article describes the development of a next-generation agent leveraging OpenAI models and the Assistant API. This agent exhibits proficiency in conducting statistical analyses of sales history datasets and generating predictions through autoregression techniques. Overall, the research underscores the promising prospects of integrating large language models into retail operations while emphasizing the imperative to broaden the scope of research in this domain. Furthermore, the application of LLMs is scrutinized through the lens of sustainability objectives, environmental implications, computational expenses, and potential risks pertaining to privacy and security. LLMs raise significant concerns regarding confidentiality and security, given the sensitivity of retail datasets, which poses risks in case of data breaches. Despite the availability of computational power, the training and deployment of LLMs remain costly. Therefore, companies must conduct comprehensive cost-benefit analyses to ensure profitability. Additionally, the substantial energy consumption associated with large computational resources raises environmental concerns. It is crucial to critically assess these impacts, even though research in this area is still in its infancy. While the adoption of artificial intelligence systems can positively contribute to sustainable development goals by overcoming existing limitations, challenges such as unequal access to technology and income redistribution need to be addressed to ensure inclusive progress.