Bank for International Settlements Innovation Hub publishes Project Spectrum research showing 34 million products can be classified for inflation analysis in five days at about EUR 1,500

The Bank for International Settlements (BIS) Innovation Hub published a Project Spectrum report on using generative AI techniques to improve inflation nowcasting by automating product classification in high-frequency price data. Built with the Deutsche Bundesbank and the European Central Bank (ECB), the approach converts product text into embeddings and then applies conventional machine learning classifiers to map items into official inflation categories. The method was tested on the ECB’s Daily Price Dataset, which contains billions of daily price-product observations for 34 million unique products, and targets classification at the European Classification of Individual Consumption by Purpose (ECOICOP) 2018 five-digit level. The report estimates that classifying the full dataset via direct GPT-5 prompting would take over six months of computing time and cost more than EUR 0.5 million, versus around five days and approximately EUR 1,500 using the embedding-based approach; average processing costs are reported at under EUR 0.031 per 1,000 products for the embedding-based classifiers compared with EUR 22.2 per 1,000 products for direct LLM prompting. On the evaluated portion of the CPI basket (around 50% coverage), direct LLM prompting achieved 86% weighted accuracy, versus 80% for a feedforward neural network and 75% for k-nearest neighbours, and the project also developed a production pipeline that can classify around one million new products in roughly three hours. Next steps highlighted include testing the solution on additional datasets and languages and running a historical back-test by constructing ECOICOP subclass indices and benchmarking them against official inflation series, which the report notes requires historical price data and is intended for a follow-up phase.