Project Spectrum：使用生成式人工智能增强通货膨胀临近预报

TitleProject GaiaTitleProject Spectrum TitleSubtitleEnabling climate risk analysis via Generative AITitleSubtitleUsing generative AI to enchanceinflation nowcasting December 2023February 2026 © Bank for International Settlements 2026.All rights reserved.Limited extracts may be reproduced or translated provided the source is stated. Follow us Executive summary The availability of web-scraped and scanner data sets provides central banks withunprecedented access to real-time data on individual product prices. However, to usethese data for inflation nowcasting and forecasting, analysts need to classify productsaccording to statistical conventions. In the absence of reliable, scalable classificationmethods, inflation analysts are flooded with data but lack actionable insight. Product classification at the scale of web-scraped data represents a major challenge.Manually processing this amount of data is not feasible. Classification using largelanguage models (LLMs) is promising, but with the LLM models currently available, theprocessing time and cost become prohibitively high. Project Spectrum used the EuropeanCentral Bank’s Daily Price Dataset (DPD), which contains billions of price-product dailyobservations for 34 million unique products. At the time of writing, classifying this dataset using GPT-5 would take over six months of computing time at a cost exceeding EUR0.5 million. Project Spectrum – a collaboration between the Bank for International Settlements(BIS), the Deutsche Bundesbank and the European Central Bank – explored an alternativeapproach where artificial intelligence (AI) was used only to transform product descriptionsinto high-dimensional text embeddings. These were then classified into product categoriesusing classic machine learning algorithms. Text embedding is a foundational AI techniqueused by many natural language processing applications. This method achieved accuracylevels comparable to LLM prompting, but at a fraction of the cost: the entire DPD wasclassified in just five days for approximately EUR 1,500. Besides classifying all records in the current DPD, the project has developed aproduction pipeline solution that can classify new products as they are added to the DPD.In addition, to ensure continuous improvement, an iterative algorithm was implementedto gradually expand the reference data set. By selectively adding manually labelled datato the reference and validation sets, this algorithm systematically refines the classificationlogic, enhances overall predictive accuracy and adapts to a changing product range. By turning raw, fragmented product descriptions into structured data, ProjectSpectrum equips analysts and policymakers with timely, detailed insights into pricedevelopments. Ultimately, the project contributes to an emerging new generation of AI-powered analysis, where data abundance can be translated more easily into actionableeconomicunderstanding. This report is intended for monetary policy analysts who utilise high-frequency datafor inflation nowcasting and data scientists within central banks looking for cost-effectivealternatives to LLMs for large-scale classification. It also serves as a technical reference forstatistical agencies seeking to automate the categorisation of scanner and web-scrapeddata into official indices. Finally, it provides a methodological framework for economicresearchers studying price-setting behaviour at the individual product level. Acronyms and abbreviations Table of contents Executive summary3 1. Using online price data for inflation nowcasting6 3. Spectrum overview9 4. Data124.1. Daily Price Data set (DPD)124.2. Ground truth via manual labelling of reference and test data sets134.3. Sample representativeness15 5. Implementation 5.1. Curated reference data set175.2. Embedding model185.3. Classification using the k-nearest neighbour algorithm195.4. Classification using a feedforward neural network205.5. Direct large language model prompting21 6. Evaluation 6.1. Classification accuracy226.2. Feasibility and cost comparison of classification methods26 7. Initial deployment and a continuous refinement process28 8. Conclusions and next steps 9. Project participants and Acknowledgements31 10. Bibliography 32 1.Using online price data for inflation nowcasting Accurate and timely inflation nowcasts1and forecasts are central to effective monetarypolicy because inflation often responds to policy measures with a time lag.2By anticipatingfuture price trends, policymakers can adjust interest rates and other policy tools earlier tomaintain price stability and prevent costly economic fluctuations. Such foresight not onlyhelps anchor inflation expectations and foster sustainable growth,3but also reinforcespublic confidence in the monetary authority’s ability to respond promptly and effectivelyto emerging risks. As a significant share of consumer spending is done online, data from e-commerceplatforms are a rich source of real-time info