What can newspaper articles reveal about the euro area economy?

Abstract
This study introduces a novel approach to dictionary-based sentiment analysis that extracts valuable insights from economic newspaper articles in the euro area without requiring article translation. We develop sentiment indices that accurately measure economic, labour, and inflation perceptions in Germany, France, Italy, and Spain using native language texts. The aggregation of these country-specific sentiments provides a reliable indicator for the euro area economy, demonstrating the effectiveness of our approach in several nowcasting and forecasting experiments. This translation-free method significantly reduces resource requirements, is simple to replicate across various languages, and enables daily updates.

Motivation and Methodology

Economic research has extensively examined sentiment shocks as drivers of business cycle fluctuations, finding that they can explain a significant part of such variations (e.g., Lagerborg et al. (2023)). Sentiment shocks refer to changes in the optimism or pessimism of agents’ expectations that are unrelated to economic fundamentals. Keynes famously described such shifts as ‘animal spirits’ or the state of confidence, considering them key determinants of investment. Since sentiment is not directly observable, proxies are required to measure it. Surveys have traditionally served as effective tools for capturing consumer and business expectations. More recently, advancements in natural language processing (NLP) have enabled sentiment extraction from textual data, such as news articles, social media posts and financial reports.

Building on this research, we construct euro area newspaper sentiment indices for output, inflation, and employment using a dictionary-based approach. A major advantage of our approach is that it eliminates the need for machine translation, thereby avoiding the financial and computational costs of translation APIs and the potential security risks of transmitting proprietary text to external services. This feature is particularly valuable in multilingual contexts, such as in our study, which analyses German, Spanish, French, and Italian newspaper articles obtained from Factiva Dow Jones.

To ensure accurate sentiment extraction, we developed language-specific dictionaries with input from native-speaking economists. The indices are constructed through the following steps:

• Article Filtering: News articles mentioning the country of origin (e.g., Germany) are filtered to focus on domestic economies. Topic-specific keywords and their variations (e.g., “econom*” for economy) are identified in each article.

• Sentiment Detection: Sentences containing these keywords are analysed within a +/-5-word window to identify directional terms indicating sentiment (e.g., “improv*” for increase, “slow*” for decrease). Single words with strong directional implications (e.g., “deflation” for price decreases) are also included.
• Sentiment Classification: Articles are classified as positive, negative, or neutral based on the presence of directional terms. An article is positive if it contains more sentences classified as ‘increase’ than ‘decrease,’ negative if the reverse is true, and neutral if both types of sentences are equally present.
• Daily Sentiment Scores: Daily sentiment scores are calculated as the difference between the number of positive and negative articles scaled by the total number of relevant articles mentioning the country. Indicators are standardized using the pre-COVID period to ensure comparable volatility across newspapers.
• Aggregation: country-specific sentiment scores are aggregated using a simple average to derive euro area sentiment indices.

This research work demonstrates that these indices provide valuable insights for forecasting GDP and inflation, while also serving as effective tools for tracking the business cycle.

Sentiment and the Business Cycle

We begin by introducing the Economic News Sentiment (ENS), a novel textual sentiment indicator derived from newspaper articles. Figure 1 shows the ENS alongside quarterly real GDP growth in the euro area and two widely used monthly survey-based sentiment measures: i) Composite Purchasing Managers’ Index for output (pmicomp), published by S&P Global, which covers both manufacturing and services sectors and is one of the most watched indicators; and ii) Economic Sentiment Indicator (ecesi), published by the European Commission, which aggregates sentiment in manufacturing, construction, retail, and services sectors and by consumers.

Figure 1. News and survey sentiment vs GDP growth

The strong co-movement between ENS and survey-based sentiment measures underscores the reliability of our textual sentiment approach. All sentiment measures are procyclical, exhibiting strong positive correlations with GDP growth. These findings are in line with prior studies (for example, Ashwin et al. (2024); Aguilar et al. (2021); Aprigliano et al. (2022)).

In addition to economic sentiment, we constructed two recession-related indices:

• Recession Word (Rword): Measures the share of economic articles that mention the term “recession”, indicating concerns about economic downturns, whether due to the risk of recession or the economy already being in recession (Ferrari Minesso et al., 2022; Bybee et al., 2024).
• Recession Risk (Rrisk): Measures the share of economic articles containing both the word “recession” and terms like “uncertainty” or “risk”, emphasizing fears of a potential recession (Hassan et al. (2023)).

Figure 2 presents the quarterly trends of daily ENS, Rword and Rrisk indices for the euro area, with the official CEPR recession dates shaded in grey. The recession indices exhibit clear countercyclical behaviour, rising during economic downturns, while ENS shows procyclical behaviour, improving during economic expansions. All three news-based indicators provided negative signals prior to the financial crisis and the sovereign debt crisis. However, the negative signal preceding the COVID-19 crisis was less pronounced, likely due to the unpredictable nature of the pandemic.

Figure 2.  Euro area news indices (90-day moving averages)

The Inflation News Sentiment (INS) provides valuable insights into inflation dynamics by capturing public sentiment from news articles. Recent studies, such as those by de Bandt et al. (2023) and Angelico et al. (2022), have shown that textual indicators can explain inflation and inflation expectations in countries like France and Italy. Building on this work, Figure 3 compares euro area INS with actual inflation (year-on-year growth of the Harmonized Index of Consumer Prices (HICP)) and 12-month ahead household inflation expectations from the European Commission consumers’ survey. INS exhibits forward-looking properties, often anticipating changes in inflation trends. For example, during the Great Recession, the INS reached its lowest point in December 2008, followed by an upward shift mirrored by actual inflation a year later. Similarly, the recent downward shift signalling the end of the high inflationary period experienced in 2022 occurred first in the INS (July 2022) than in the actual inflation (March 2023). While INS demonstrates predictive capabilities for inflation trends, its relationship with household inflation expectations is particularly strong contemporaneously, with a correlation of 0.76. This is consistent with Carroll (2003)’s theoretical framework, which posits that individuals frequently revise their inflation expectations based on the latest information available in the news.

Figure 3. Inflation, News Sentiment and Household Expectations

Are News Useful to Predict the Euro Area Economy?

We investigated the effectiveness of news-based sentiment indices as reliable tools for forecasting economic trends in the euro area. Our findings show that the ENS and Recession indices outperform traditional indicators, such as industrial production, retail sales, Purchasing Managers’ Index, and the Economic Sentiment Indicator, in quarterly GDP short-term forecasting. In a daily real-time nowcast exercise, incorporating news sentiment indices into forecasting models improved ECB benchmark forecasts, especially during the early part of the quarter when key macroeconomic data, such as industrial production and retail sales, are unavailable.

Our indices, computed directly from the non-translated articles, demonstrated better nowcast performance throughout the quarter when compared to the news sentiment indices derived from English-translated articles by Ashwin et al. (2024). This also demonstrates that working directly with non-translated articles can enhance forecast accuracy while reducing computational complexity and costs.

The INS index proved highly effective in predicting the inflation spike of 2023 caused by the energy supply shortage, outperforming both consensus forecasts and household inflation expectations.

While prior studies have utilized newspaper sentiment to analyse economic growth in individual countries, few have exploited its high-frequency availability. To address this gap, we show that daily newspaper sentiment can provide recession probabilities with high accuracy. Using a simple probit model to predict CEPR euro area recession dates based on quarterly moving averages of daily sentiment data as predictors achieved excellent out-of-sample forecasting performance, with an area under the ROC curve (AUC) of 0.93 when tested on the sovereign debt and COVID-19 pandemic crises. Although some false recession signals occurred during 2019 and 2022, these were short-lived and less pronounced compared to signals during actual recessions. In our sample, when newspapers reported concerns about a recession, it was highly likely that the economy was indeed in a recession. Even in cases of overestimated recession risks, these indicators remain valuable as early warnings for policymakers, highlighting potentially negative sentiment trends.

Conclusions

The increasing availability of textual data offers a unique opportunity to construct sentiment indicators that complement traditional macroeconomic and survey-based measures. Motivated by the need for timely and reliable economic insights, particularly in the aftermath of the COVID-19 pandemic, this study evaluates the value of sentiment indicators derived from news articles. Building on previous research (e.g., Aguilar et al., 2021; Aprigliano et al., 2022; Barbaglia et al., 2024; Ashwin et al., 2024), we demonstrate that newspaper sentiment effectively tracks inflation and output, predicts turning points in the euro area economy, and provides valuable forecasting capabilities.

Our findings reveal that newspaper sentiment outperforms traditional indicators in GDP forecasting and is particularly useful for nowcasting during the early part of the quarter when conventional macroeconomic data, such as industrial production and retail sales, are not yet available. By leveraging Factiva’s daily database updates, these news-based sentiment indicators also provide accurate daily recession probabilities in the euro area. Additionally, inflation-related newspaper sentiment closely tracks household inflation expectations and anticipates inflation trends, proving its forward-looking utility during the 2022–2023 inflationary period triggered by the Russia-Ukraine conflict.

Furthermore, translating the articles into English, a common practice in the literature, was found to be unnecessary for producing high-quality sentiment indices in a multilingual framework. When comparing our approach for building news-based sentiment indices with other dictionary-based methods previously used for English texts, our methodology yielded lower forecast errors in GDP short-term forecasting.

Finally, the textual analysis method used to construct these indices is both straightforward and efficient. By relying on small, language-specific dictionaries with few words that can be easily adapted to other languages, the approach eliminates the need to translate entire texts, significantly reducing computational costs and simplifying automation. Furthermore, avoiding machine learning techniques ensures consistency, interpretability, and transparency, as the results are not dependent on a training sample, and the inner workings of the algorithms are easy to understand. The simplicity and reliability of this methodology enhance the replicability of these sentiment indices, making them practical tools for economic analysis.