Bank for International Settlements research finds word embeddings from European Central Bank press statements can improve core inflation forecasts

The Bank for International Settlements published a working paper assessing whether word embeddings extracted from central bank communications can help predict inflation. Using the European Central Bank’s introductory statements at monetary policy press conferences as a case study, the paper finds that embedding-based text measures improve out-of-sample forecasts of euro area core inflation multiple quarters ahead and outperform common sentiment and dictionary-based approaches. The authors estimate Word2Vec embeddings in a real-time recursive framework using only information available as of each press conference, summarise the resulting high-dimensional vectors with the first four principal components, and include these factors in a Bayesian vector autoregression with core inflation (Harmonised Index of Consumer Prices excluding energy and food). Over 2008–2023, the embedding-augmented model beats an autoregressive benchmark at horizons of one to four quarters, with the four-quarter-ahead mean squared forecast error more than 16% lower; placebo measures such as the count of the word “inflation” and statement length worsen forecasting performance, while sentiment indicators deliver smaller gains. Forecasts based on pre-trained embeddings from BERT and OpenAI perform better in the full sample but may reflect look-ahead bias; in the pre-COVID 2008–2019 subsample, Word2Vec is competitive and performs at least as well at longer horizons. An encompassing exercise suggests the text-based signal is not fully captured by the Eurosystem’s macroeconomic projection exercise, particularly at longer horizons. The paper flags open questions around interpreting the embedding factors, handling uncertainty in two-stage text-to-forecast procedures, and testing whether results generalise across central banks with different communication practices.