Abstract

This study assembles a novel high-frequency dataset of product-level web-scraped supermarket prices to improve the nowcasting of food inflation in Malta. The dataset comprises more than 2.2 million daily price observations across over 2,700 food products sold by two major supermarket chains and a network of corner-shops. Products are systematically cleaned and classified using a hybrid approach that combines string-matching methods with large language model (LLM)–assisted categorisation. Using this dataset, three nowcasting approaches are implemented: (i) a naïve benchmark, (ii) a minimum-distance methodology aligned with official National Statistics Office (NSO) micro-category price indices, and (iii) a machine learning (ML) framework incorporating mixed-frequency regressions. The results show that web-scraped prices yield meaningful forecasting gains relative to established benchmarks, particularly for categories with stable product coverage and rich high-frequency variation. However, predictive performance relative to the Narrow Inflation Projection Exercise (NIPE) varies between categories, reflecting heterogeneity in pricing behaviour and product properties. The findings demonstrate the feasibility and value of integrating high-frequency online data with econometric and ML frameworks in a small open economy.