From Words to Forecasts: ECB Leverages NLP to Predict Euro Area Inflation Trends

The ECB’s Word2Prices study shows that word embeddings derived from central bank statements can significantly improve inflation forecasts. Using a simple model like Word2Vec, the research captures meaningful insights from policy language beyond traditional sentiment or projections.

In a groundbreaking study titled “Word2Prices: Embedding Central Bank Communications for Inflation Prediction,” researchers from the European Central Bank (ECB), the Bank for International Settlements (BIS), and the European Stability Mechanism (ESM) unveil an innovative approach to macroeconomic forecasting that leverages natural language processing (NLP) to decode central bank language. By applying machine learning techniques to the ECB’s press conference statements, the authors show that central bank communication is not just narrative—it’s data. Their findings reveal that these carefully crafted texts contain predictive signals that can significantly enhance inflation forecasting, especially when traditional models fall short.

Turning Words into Numbers

At the heart of the research is the concept of word embeddings—mathematical representations of words in a vector space that encode semantic meaning based on contextual usage. The authors employ Word2Vec, a relatively simple but powerful NLP model, to process the ECB’s introductory statements delivered during monetary policy press conferences. By capturing how words co-occur in sentences, Word2Vec generates numerical representations for each word. These word vectors are then averaged first across individual speeches and then across each quarter to create a quarterly time series reflecting the central bank’s messaging. This metric is designed to capture subtle shifts in tone, emphasis, and narrative framing over time.

Unlike modern large language models (LLMs) like BERT or OpenAI’s GPT, which require vast computational resources and often rely on pre-trained embeddings that can introduce look-ahead bias, Word2Vec can be retrained from scratch using only historical data. This makes it ideal for a rigorous out-of-sample analysis that mimics the real-time constraints faced by economists and policymakers. The ability to update the model with only the data available up to each point in time ensures methodological integrity and avoids the pitfalls of hindsight-driven results.

Forecasting Inflation with Text

The research focuses on core inflation in the euro area, using the Harmonised Index of Consumer Prices excluding food and energy (HICPex) as the primary variable. The authors integrate the word embedding time series into a Bayesian Vector Autoregressive (VAR) model and compare its forecasting performance against standard autoregressive (AR) benchmarks. The results are striking: the VAR model augmented with text-based variables consistently outperforms the AR model across forecast horizons of one to four quarters. The advantage is especially notable at longer horizons, where traditional models often struggle due to inflation’s inherent volatility and external shocks.

Even when compared with forecasts based on embeddings from more sophisticated LLMs like BERT and OpenAI, the Word2Vec-based model remains highly competitive. While the more advanced models show slightly better performance over the full sample, their superiority diminishes significantly when the pre-COVID period is considered in isolation. In fact, Word2Vec outperforms them at longer forecast horizons during that more stable period. This reinforces the practical value of a simple, adaptable model in institutional settings where real-time updates and computational efficiency are key.

Beyond Sentiment: The Power of Embeddings

To validate the robustness of their approach, the authors compare Word2Vec embeddings with several alternative text-derived indicators. These include basic sentiment scores using dictionary-based methods, word counts (e.g., frequency of the word “inflation”), and the overall length of the statements. These "placebo" metrics either had a negligible impact or worsened the model’s predictive performance. Sentiment metrics fared slightly better but still fell short of the embedding-based approach, indicating that traditional sentiment analysis fails to capture the depth of information encoded in central bank communication.

The success of the embeddings lies in their ability to pick up not just what is said, but how it is said, the structure, word choice, and context that reflect evolving policy priorities and economic outlooks. Embeddings translate this nuance into quantifiable data that can meaningfully augment statistical models. It’s not merely about detecting whether a statement is “hawkish” or “dovish,” but rather about capturing the subtle evolution of narrative emphasis across time.

Complementing Official Forecasts

An intriguing question the paper addresses is whether the predictive power of text embeddings merely replicates the official forecasts made by the ECB’s staff, known as the Broad Macroeconomic Projection Exercise (BMPE). The authors show that this is not the case. Through optimal forecast combination tests, they find that embedding-based forecasts provide unique information not already captured in BMPE projections. This is especially true at longer horizons, where the text-based forecasts command substantial weight in the combined prediction.

This result suggests that the ECB’s language contains layers of meaning beyond the raw numerical projections. It may reflect the Governing Council’s risk assessments, judgment calls, and policy nuances that are not fully embedded in the formal forecasts. Thus, analyzing central bank language not only improves forecasting but also offers a lens into the institution’s internal thinking.

Bridging Text and Economics

The implications of Word2Prices are far-reaching. It provides a low-cost, easily implementable method to extract valuable economic signals from central bank texts. The approach has immediate applications for central banks, financial institutions, and policy analysts seeking to enhance forecasting accuracy without relying on black-box models. It also lays the groundwork for further exploration into how different communication styles across central banks may affect economic expectations and market behavior.

In an age where machine learning and textual data are reshaping disciplines, this paper is a timely reminder that language, when treated as data, can offer powerful insights. With methodological rigor and practical relevance, the authors demonstrate that what central banks say and how they say it can indeed help predict where inflation is heading.