BUILDING A HIGH-FREQUENCY MEDIA INDICATOR OF REGIONAL ECONOMIC ACTIVITY BASED ON ANALYSIS OF MEDIA PUBLICATIONS USING LARGE LANGUAGE MODELS

Abstract

The inherent publication lags in official statistical releases have precipitated a growing scholarly and practical interest in alternative data sources for economic assessment. Among these, the vast corpus of textual data from media outlets, processed through Large Language Models (LLMs), presents a significant opportunity. This study aims to develop a media-based indicator of economic activity for a constituent entity of the Russian Federation by analyzing regional news publications with LLMs. The empirical foundation of this research comprises a corpus of texts from online media spanning the period 2021–2025. The methodological framework consists of several sequential stages: (i) the compilation and preprocessing of a regional news publication database; (ii) the thematic classification of texts according to aggregated economic sectors employing a zero-shot classification paradigm; (iii) the extraction of economic sentiment from the news articles using the GigaChat large language model; and (iv) the construction of an aggregated media indicator. The resulting output is an integral media indicator of economic dynamics, designated as the IBVEDMedia index. A comparative analysis with the official Index of Output of Goods and Services for Basic Economic Activities reveals a statistically significant correlation between the two time series. This finding suggests that the news flow from regional media sources encapsulates substantial information pertinent to the current state of the regional economy. The proposed media indicator can be considered a composite economic sentiment indicator, serving as a valuable complement to traditional official statistics. The validation of its real-time operational properties is a subject for future research.