Economic Sentiment Analysis Based on 200 Years of News Text

Although many scholars have studied measures of economic sentiment in recent years, their application has been somewhat limited due to short time series and insufficient granularity. This paper uses machine learning methods to construct a new measure of economic sentiment based on the published pages of local newspapers in the US over the past 170 years. We demonstrate that economic sentiment is crucial for understanding business cycles at both local and national levels. We find that economic sentiment can predict economic fundamentals at both national and state levels, leading common GDP forecasts; there are significant cross-sectional differences across states, with national factors accounting for only 35% of state differences.

To achieve this, we utilized a historical collection of digitized newspapers over 170 years, which includes text from 200 million pages of 13,000 local newspapers. Our corpus comprises approximately 1 billion newspaper articles, representing a significant increase compared to the Wall Street Journal corpus, which is a commonly used text data source in economics and finance, containing about 1 million articles. In fact, our data is about 95 times the total number of English Wikipedia entries. By leveraging the text information from local newspapers, we can measure sentiment at a higher granularity (e.g., at the county or state level).

To measure text-based economic sentiment, we employed the machine learning techniques pioneered by Singla and Mukhopadhyay (2022). We created a fully automated, topic-specific dictionary using a neural network-based algorithm called Word2vec, which enables us to capture the meanings of words and phrases from their contexts. According to Hamilton, Clark, Leskovec, and Jurafsky (2016), we did not use binary positive/negative labels but generated a continuous sentiment measure for each word and phrase in the dictionary. Thus, our method automatically overcomes many common challenges faced by simple word counting techniques, such as detecting negations and measuring word/phrase intensity.

Our measurement method shows significant differences in both time series and cross-section. Furthermore, our method reveals that news sentiment has exhibited a significant downward trend since the 1970s. We show that the reporting of economic news has not driven this trend; rather, it involves general reporting of both economic and non-economic news. By stripping this overall sentiment from economic news, we constructed an economic sentiment indicator that is unaffected by changes in the overall tone of newspapers.

Consistent with recent advances in natural language processing, we measure sentiment as the positivity versus negativity of economic news coverage in the mass media. Therefore, our measure can reflect the pure factual reporting of economic news, irrational exuberance, or some combination of both. In terms of recent asset pricing papers, our measure can reflect changes in cash flow expectations and discount rates. In this sense, the term “sentiment” in our paper relates more to the general meaning of the English word rather than the more specific interpretations it has received in behavioral economics and finance recently.

Even after controlling for current GDP growth, our indicator can still predict future per capita GDP growth: during the sample period from 1850 to 2017, a one standard deviation increase in the sentiment indicator corresponds to an additional 2% GDP growth in the following year. Moreover, during the sample period from 1947Q1 to 2019Q4, where the slope of the yield curve (a standard recession forecasting factor) can be observed, we found that a one standard deviation increase in the sentiment indicator leads to an additional 0.29% GDP growth in the next quarter (equivalent to an annualized growth of 1.1%). Furthermore, this predictability persists even after controlling for the consensus forecasts in professional forecaster surveys, indicating that our measure captures important information not covered by such consensus forecasts. Notably, as we demonstrate, our indicator is actually leading in the survey, as it predicts the consensus value for the next quarter. Overall, our findings illustrate the importance of sentiment in understanding business cycles.

The flexibility of our method allows us to further explore the sources of predictability in our measure. Articles in newspapers include both current affairs reporting and future outlooks and opinions. To distinguish these two types of news, we created specific topic dictionaries related to current and future events and calculated the intensity of this news language on each newspaper page, categorizing it as either current-related or future-related. We used this classification to create current and future economic sentiment. We found that all of our predictability results are driven by the forward-looking, future component of sentiment, while indicators created based on current events have little predictive power over macroeconomic variables. This suggests that our indicator reflects forward-looking information about economic fundamentals.

Given that our indicator predicts GDP growth, one might wonder which inputs to GDP our economic sentiment indicator predicts. We show that our indicator primarily operates through the labor channel rather than the capital channel. It predicts employment, consumption, and services, but does not predict investment and industrial production. Similarly, we demonstrate that our measure of economic sentiment is related to the real economy rather than predicting inflation.

Next, we assess the extent to which news-based economic sentiment is reflected in daily policy decisions. To this end, we quantified the importance of sentiment in explaining changes in the federal funds rate relative to the forward-looking Taylor rule framework proposed by Romer and Romer (2004). We found that sentiment has a significant impact on key policy rates: a one standard deviation decline in the sentiment indicator over the past two quarters leads to a 25 basis point decrease in policy rates during a recession. Furthermore, we found that even after controlling for the predictive ability of sentiment for pre- and post-recession GDP growth, sentiment still has significant predictive power for the federal funds rate during recessions.

Our method also allows us to measure sentiment at a more granular level, revealing significant cross-sectional heterogeneity among states. Common factors among states account for only about 35% of the changes in sentiment at the state level. Even after controlling for national-level sentiment indicators and national and state GDP growth, state-level sentiment can still predict state GDP growth. Additionally, using the dispersion of sentiment across states as a measure of heterogeneity, we found that, after controlling for national-level sentiment indicators and current national GDP growth, higher dispersion predicts lower future GDP growth levels in the nation.

Sentiment can manifest through two channels: the discount rate channel and the (expected) cash flow channel. The former operates through financial markets, where higher company valuations are accompanied by lower discount rates (thus lowering future returns), stimulating corporate hiring and investment decisions. The latter requires sustained changes in discount rates, not just short-term changes caused by sentiment. In contrast, the second channel operates solely through cash flows and their expectations, and thus does not require changes in discount rates. Higher sentiment regarding future economic activity stimulates current hiring and (potential) investment, which can justify the improvement in current sentiment.

Compared to the contributions mentioned above, our economic sentiment indicator does not have much predictive power for variables such as stock returns. Instead, as reported in newspaper articles and reflected in our measure, forward-looking economic sentiment has significant predictive power for future fundamentals such as GDP and labor. Therefore, our paper does not provide direct evidence supporting behavioral biases that stimulate real economic activity through exaggerated asset valuations (lower discount rates), leading to boom-bust cycles. That is to say, our paper does not rule out the view that sentiment can act as a self-fulfilling prophecy (rational sentiment) that operates through the cash flow channel: anticipated fluctuations can lead to hiring cycles, which in turn lead to fluctuations in GDP growth. Given that we find minimal impact on corporate investment, our findings are most consistent with short-term sentiment fluctuations stimulating short-term hiring.

While the higher granularity and longer historical coverage of text data allow for the construction of more useful metrics, the relevant increase in granularity and time span also presents challenges. Natural language processing techniques require large text data corpora, typically at a level capable of constructing metrics. One contribution of our paper is to demonstrate how to create new state, county, or city metrics using text data. Another challenge arises from the fact that the context and meaning of words change over time. Proper historical text analysis requires a dictionary that can capture the historical meanings and contexts of relevant words. In this paper, we applied a method capable of constructing an automatic dictionary to address this issue.

The remainder of this paper is organized as follows. The first section describes the coverage and granularity of the newspaper text data we used to construct measures of economic and non-economic sentiment. The second section introduces the various components of our method. The third section presents the time series and cross-sectional dynamics of national and state-level economic and non-economic sentiment indicators. The fourth section highlights the relationships between national and state economic sentiment, economic activity, and business cycles. The fifth section explores various aspects of the interaction between economic sentiment and monetary policy. The final section concludes.

A key advantage of this paper is the size and granularity of the text we used to construct measures of economic and non-economic sentiment. We utilized the historical collection of digitized newspapers introduced for the first time in Singla and Mukhopadhyay (2022) over 170 years. The data includes the full text of 200 million pages from 13,000 local newspapers, equivalent to about 1 billion newspaper articles.

Sentiment analysis in economics and finance typically follows three broad steps: first, creating a relevant word dictionary for a given topic, which includes economic-related terms and phrases in our setup. Second, identifying words that convey positive or negative sentiments, such as “growth” versus “recession.” Finally, using the dictionary to measure the intensity of positive and negative sentiment in specific documents (e.g., articles or newspaper pages). We followed these common steps and implemented them using various machine learning techniques at each stage.

To construct sentiment measures, we introduced a natural language processing technique called word embedding, which uses the co-occurrence of words to produce vector representations of these terms and phrases. This process captures the semantic information relevant to each term or phrase based on its context. The objective function of the algorithm ensures that similar words have similar vector representations, meaning they have similar meanings. Note that Word2vec is applicable to phrases and individual words and is known to outperform existing alternatives. Once the algorithm produces these vector representations, we can measure their similarity (cosine similarity) based on the distance between them. Because the algorithm focuses on the context of word/phrase usage rather than the physical distance between words in a given sentence, it can identify similar terms that do not appear together as long as the adjacent contextual words and phrases are sufficiently similar.

In contrast to traditional word counting methods in economics and finance that treat words as distinct objects, word embedding allows us to infer relationships between words/phrases by directly studying their contexts. Therefore, although still not frequently used, word embedding techniques have recently become the frontier of NLP-based applications in the economics and finance literature.

In the first step of our method, we used Word2vec to create a fully automated, topic-specific dictionary of economic-related words/phrases. To train our word vectors, we employed the skip-gram implementation of Word2vec. Two key hyperparameters are crucial for training the Word2vec model: the dimensionality of the vector embeddings and the window length that determines each word’s neighborhood. Previous research has shown that a vector dimension size of 300 is sufficient to effectively capture word meanings in downstream NLP tasks. Therefore, we set the vector dimension to 300. Following the specifications used in the application of Word2vec in economic literature, we used a standard window length of 10.

In the sample from 1850 to 2020, we randomly selected 1 billion words every five years to ensure that the corpus is historically balanced and that each time period has a consistent representation. Based on the vectors trained by Word2vec, we generated a dictionary of words and phrases related to economics. Among them, the term “Economy” is most closely associated with the following terms and phrases: “Economic growth,” “inflation,” “economic,” “recession,” “economic recovery,” “consumer spending,” etc., which also preliminarily validates the effectiveness of our method. This automated dictionary is characterized by its size and diversity, calculating the top 1000 words and phrases most strongly correlated with the term “Economy,” as shown in the word cloud in Figure 1.

Figure 1 Word cloud of the top 1000 words most similar to “Economy”

Of course, a standalone economic-related vocabulary dictionary is insufficient for measuring economic sentiment. To reflect the latter, each word in our dictionary must be classified as either positive or negative. Therefore, we referred to the literature on dictionary generation using word vectors. Sentprop is a label propagation algorithm that begins with several initial seeds and classifies all words in the dictionary. Another advantage of the algorithm is that it can generate continuous polarity scores for each word in the corpus. Thus, it can reflect not only the positive or negative connotation of a word or phrase but also the strength of its polarity. This method achieves state-of-the-art performance in accurately measuring sentiment in economic and financial contexts.

In this section, we propose a new text-based measure of economic sentiment. We introduce a national-level measure and a state-level measure in sections III.1 and III.2, respectively.

After establishing the time series and cross-sectional characteristics of national and local sentiment measures, we now analyze their informational content. In the following subsections, we combine traditional economic growth forecasting models with economic sentiment and examine whether existing macroeconomic indicators capture the dynamics and potential predictive impacts of the latter.

This section explores three aspects of the interaction between economic sentiment and monetary policy. First, we examine whether central banks consider economic sentiment in the interest rate-setting process. Second, we quantify the degree to which economic sentiment influences decision-making compared to other macroeconomic variables. Finally, we explore the pathways through which economic sentiment affects monetary policy.

Utilizing 193 million pages of articles from local US newspapers and state-of-the-art machine learning methods, we constructed economic sentiment indicators at both state and national levels over nearly two centuries. During the sample period in which both indicators exist (1978–2017), our indicators are highly correlated with existing survey-based economic sentiment indicators. Our measures significantly expand the temporal availability and granularity of measuring economic sentiment. During the major economic recessions in the US from the 19th to the 21st century, sentiment exhibited significant time series changes and substantial declines. We demonstrate that even after controlling for the macroeconomic information available at the time of prediction, our indicators can predict economic fundamentals such as GDP, consumption, and employment growth.

Sentiment also exhibits significant cross-sectional differences among states, with common components accounting for only 35% of the total differences. Even after controlling for national-level sentiment and GDP growth, state-level sentiment predicts local GDP growth. Furthermore, cross-sectional differences in state sentiment are associated with declines in national GDP growth rates. In summary, our findings indicate that sentiment at both local and national levels plays a crucial role in understanding business cycles.

……

Get the Full Report

Please reply “Economic Sentiment” in the background

to obtain the download link

Editor/Song Qingxin

Proofreader/Qin Sihan

Editor-in-chief/Yang Shiqi