European Central Bank working paper finds word-embedding analysis of policy press statements improves core inflation forecasts

The European Central Bank has published Working Paper No 3047 examining whether word embeddings extracted from the ECB President’s introductory statements at monetary policy press conferences can be used to predict euro area core inflation. The authors report that embedding-based text measures improve out-of-sample forecasting performance multiple quarters ahead versus standard autoregressive benchmarks, while commonly used dictionary and sentiment-style text metrics deliver smaller gains. The paper estimates Word2Vec embeddings using only information available at each press conference and re-estimates the model each quarter to support a real-time, out-of-sample design, then summarizes the embedding vectors using the first four principal components within a Bayesian vector autoregression alongside core inflation (Harmonized Index of Consumer Prices excluding energy and food). Over 2008–2023, the four-quarter-ahead mean squared forecast error is reported to be more than 16% lower than the autoregressive benchmark for the Word2Vec-based approach, while embeddings derived from pre-trained BERT and OpenAI models perform even better but may be affected by look-ahead bias. An encompassing-style forecast combination exercise suggests the text-based forecasts are not fully redundant with Eurosystem projection information, with non-negligible optimal weights on the text-based forecast (for example, around 0.37–0.56 across one- to four-quarter horizons in the pre-COVID sample, and 0.68 at the four-quarter horizon in the full sample).