Project Spectrum: using generative AI to enhance inflation nowcasting

19 November 2024

Inflation forecasters have an abundance of data at their fingertips and are continually developing new models to distil actionable insights from this wealth of data. However, available data on consumer prices is often raw, high-volume and too unstructured to be processed manually: using it for inflation nowcasting requires extensive automation.

The Bank for International Settlements (BIS) is launching Project Spectrum to explore how generative artificial intelligence (AI) can be used to automatically categorise billions of individual product descriptions and price observations to improve inflation nowcasting. Project Spectrum is an initiative of the BIS Innovation Hub Eurosystem Centre with the European Central Bank and the Deutsche Bundesbank.

It aims to convert raw big data into accessible information using state-of-the-art AI technology to enable accurate nowcasting of inflation. The project leverages the European Central Banks's Daily Price Dataset, the Eurosystem Centre's AI learnings from Project Gaia and the multi-lingual abilities of large language models (LLMs). Spectrum sets out to prove the effectiveness of LLMs for structuring big data, using European data as a first step. In the initial experiments the algorithms will map the data to globally standardised household expenditure categories such as the United Nations' Classification of Individual Consumption According to Purpose (COICOP).

The developed algorithms and insights could be applicable beyond Europe: the training dataset covers global retail chains and multiple European languages, and the LLMs and embedding technologies allow for the findings to be extrapolatable even to languages not included in the training set.