NewsflationFrançais

How the indicator is built

From raw press articles to an official monthly series, in six steps.

  1. Collect

    More than one million articles from 30 French national and regional newspapers, magazines and news agencies.

  2. Select

    Keep the articles that use the vocabulary of prices and inflation.

  3. Filter

    Remove other meanings of the same words, such as literary prizes, so only consumer prices remain.

  4. Classify

    A supervised machine-learning model labels each article: prices rising, falling or stable.

  5. Aggregate

    A monthly balance of opinions, plus the number of articles as a measure of media attention.

  6. Validate

    Compared with the European Commission household survey and with HICP inflation.

1. Corpus

The corpus holds more than one million articles published since 2004 by 30 French sources: national and regional daily newspapers, weekly magazines and news agencies, in print and online. Articles are accessed under licence. No article text is redistributed on this site, only monthly aggregates.

2. Selecting articles about prices

Articles are first retrieved with keywords from the vocabulary of prices and inflation. Many of these words have other meanings: in French, prix is both a price and a prize. A filtering step keeps only the articles that are about the prices of goods and services.

3. Classifying the direction of prices

A supervised machine-learning model, trained on articles labelled by hand, assigns each article a direction: prices rising, falling or stable. This approach is compared with alternatives based on topic models and on word pairs, and it performs better on French data.

4. Two monthly indicators

The first indicator counts the articles about prices each month, as a measure of media attention to inflation. The second is a balance of opinions, built like the balances used in household and business surveys: articles reporting rising prices against articles reporting falling prices. The published series comes in a raw version and a version smoothed with an exponentially weighted moving average (α = 0.35).

5. Validation

The indicator is compared with households' inflation expectations in the European Commission consumer survey and with headline HICP inflation, and tested for its contribution to short-term inflation forecasts alongside professional forecasts, market-based measures and oil prices. The detailed results are in the paper.

6. Limitations

News coverage reflects editorial choices as well as economic conditions. The indicator measures what the press reports about prices, which shapes and reflects household views; it is not a direct measure of expected inflation. The set of sources also changes over time, which can move the number of articles independently of inflation.

Series published by the Banque de France on Webstat.