Is Perception Reality? Evidence from Cross-Language Media Sentiments on the Chinese Economy

By Marialuz Moreno Badia and Vikram Dixit Kumaraswamy

Economists have increasingly turned to news coverage to gauge economic sentiment in real time. However, this assumes there is one economic narrative to capture. For China, the world’s second largest economy, the world may not be watching through the same lens, as Chinese- and English-language media can portray the same realities very differently, shaping how financial markets and policymakers respond to economic developments.

Over the past decade, advances in textual analysis have fueled a growing body of research on news-based indicators used to track economic developments in real time and at a fraction of the cost of traditional surveys.

A new study from the Boston University Global Development Policy Center (GDP Center) built two new monthly measures of economic sentiment covering nearly two decades of media coverage of the Chinese economy: one based on English-language news, and one based on Chinese-language news. The results of the findings found a persistent and growing gap in how the media portrays the Chinese economy. The study then asked a simple but under-explored question: when English- and Chinese-language media tell different stories about the same economy, does it matter for financial markets or policymakers?

How the indices were built

Most of the other indicators used previously in this field of study share two limitations. Many rely on dictionaries that struggle to capture the context around language and produce noisy, inaccurate readings. And they are almost always built from a single language or media environment, overlooking the fact that the same economy may be portrayed differently across language-specific information environments. This second limitation is especially consequential for an economy as large as China’s.

In this case, to capture sentiment separately by language, the GDP Center study built two custom classifiers — one trained on English-language text and one on Chinese-language text — using a language model pretrained on text spanning 100 languages. Unlike a dictionary that simply counts positive and negative words, a model like this weighs context, distinguishing “growth fell short of expectations” from a straightforwardly negative statement about the economy.

Experts designed a detailed coding protocol on a random sample of 4,000 English-language and 8,000 Chinese-language articles — the larger Chinese sample reflects a lower share of economically relevant articles in that corpus and a class imbalance (Chinese media is typically more positive) — drawn from more than 1.2 million media articles, covering 2002 through 2020. These models substantially outperformed standard dictionary-based approaches according to a standard measure of classification accuracy applied across all sentiment categories.

A natural question is why not just use an AI large language model (LLM). In the study this was tested by benchmarking the classifiers against three frontier LLMs — Claude Sonnet 4, GPT-4o, and Gemini 2.0 Flash — each given the same coding instructions as our human annotators and asked to label the same held-out test articles.

The results were nuanced. On English-language articles, the LLMs performed comparably to the new fine-tuned classifier; two of the three scored marginally higher on accuracy. On Chinese-language articles, however, the study’s purpose-built model outperformed all three, particularly at distinguishing genuinely neutral coverage from coverage with a clear lean.

The results: A persistent, and widening, perception gap

Using the trained classifiers, every relevant article in the corpus was scored and aggregated into a monthly sentiment index for each language, covering August 2002 through December 2020.

Figure 1: Media Sentiment Indices and the Perception Gap, 2002–2020

The two indices move together around major shared events — the 2008 global financial crisis, the 2015 “currency fixing” episode, and the onset of COVID-19. But their average levels differ markedly. Chinese-language coverage is consistently far more optimistic: the average score is 0.39, compared with just 0.06 for English-language coverage (on an index from −1 to +1). The Chinese-language index sits above the English-language index in all but one month of the 18-year sample.

This difference can be identified as a “perception gap” — averaging about −0.33 index points, or roughly two-and-a-half standard deviations of the English-language index’s typical monthly movement. The gap is not a fixed offset: it has widened steadily over time and is also persistent — a divergence today tends to carry forward into the following months, suggesting it reflects slow-moving structural differences in how each media environment covers China’s economy rather than transitory noise.

The perception gap is striking on its own, but to explore whether this matters in real economic contexts, the study examined how movements in financial markets and policy variables are associated with innovations in each index, holding other factors constant. The answer turns out to depend entirely on who is watching.

For financial markets, one signal dominates. Market prices respond almost exclusively to the English-language index — even though Chinese-language coverage is, on average, considerably more optimistic. Once the two indices are modeled jointly, market-price movements are associated primarily with innovations in English-language sentiment; Chinese-language sentiment does not emerge as an independent systematic pricing signal.

However, the picture looks very different for policy variables. Movements in both indices are followed by statistically significant changes in money supply growth and short-term interest rates. More striking is what happens when both languages move in the same direction simultaneously: the contractionary response of money supply growth more than doubles — and the interest rate response increases by a similarly large margin. Shared narratives, it appears, are heard more loudly than divergent ones.

Why this matters

Much empirical work in international macroeconomics treats information about a country’s economy as a single signal, observed and interpreted the same way by everyone watching. Our results suggest that, for an economy as large and as differently observed as China’s, this is not quite right. English- and Chinese-language media tell related but distinct stories, and those stories are not weighted the same way by markets and by policymakers.

For market participants, English-language coverage is the dominant sentiment signal for price movements, even though Chinese-language coverage is, on average, considerably more positive. For policymakers, both narratives matter — and that relationship is strongest when the two align.

More broadly, the results caution against treating the narrative surrounding a large, globally important economy as homogeneous. These findings point to a broader shift in how information should be understood in international macroeconomics. Understanding how these parallel and differing narratives shape markets and policy could influence how researchers interpret market movements and better understand policymaking in the global economy.

*