Real-Time Inflation Expectations in China: An Explainable Nowcasting Framework and a Reproducible Communication–Attention Pilot

Real-Time Inflation Expectations in China: An Explainable Nowcasting Framework and a Reproducible Communication–Attention Pilot | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Real-Time Inflation Expectations in China: An Explainable Nowcasting Framework and a Reproducible Communication–Attention Pilot XiaoXi Ma, Lu Chao, Shulun Hou, Wang MuYao, SaiChen Jing This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9395947/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This paper develops an explainable framework for measuring real-time inflation expectations in China and embedding those signals in inflation nowcasting. The production design is LLM-ready: official texts, news narratives, and search data can be mapped into an economically structured taxonomy of inflation mechanisms. For transparent evaluation, however, the article implements a reduced-form public-data pilot rather than claiming a fully historical LLM deployment. The pilot combines monthly official CPI, PPI, M2 growth, and one-year LPR data for 2023M1–2025M12 with two reproducible expectation channels: a policy-communication index derived from PBOC materials and a lagged attention proxy that approximates public information demand under open-data constraints. In expanding-window nowcasts, expectation-augmented models improve on stricter macro-only linear benchmarks; the best full random-forest specification records an RMSE of 0.253, compared with 0.273 for a macro elastic-net and 0.359 for a simple macro linear model. Feature attribution shows that lagged CPI, PPI, and money growth remain the main anchors, while communication and the lagged attention proxy provide non-trivial incremental information. An exploratory policy-window exercise suggests that the same narrative environment tends to coincide with easing episodes, although that evidence is descriptive because the sample contains only five LPR cuts. The paper’s main contribution is therefore methodological and institutional: it shows how an LLM-ready narrative measurement system can be disciplined by transparent public data, explainable modeling, and cautious validation in China’s recent low-inflation regime. Other Business inflation expectations China nowcasting large language models explainable artificial intelligence monetary policy communication Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1 Introduction China is an especially informative setting for studying real-time inflation expectations in an open economy. During 2023–2025, the central policy problem was not persistent overheating, but rather whether weak domestic demand, a prolonged property-sector adjustment, and continued producer-price disinflation would keep price recovery too soft for too long. In a large trading economy such as China, that domestic weakness interacts with external cost, exchange-rate, and trade-policy narratives. Imported inflation can arrive through commodities, tariffs, and exchange-rate pass-through, while a weak global pricing environment can reinforce domestic disinflation. The IMF’s 2025 Article IV consultation reported that headline inflation averaged 0 percent in 2025 even though GDP growth reached the official target, and the National Bureau of Statistics reported that 2025 CPI was flat on average while producer prices fell 2.6 percent over the year. In that environment, expectations matter because policymakers must prevent a low-inflation equilibrium from becoming self-reinforcing and must judge how domestic and external narratives are affecting price formation in real time. That policy problem immediately creates a measurement problem. Inflation expectations move faster than official price releases. They respond to food-price headlines, policy-support narratives, property-market news, tariff and exchange-rate stories, and the evolving tone of official communication. In China, these channels are especially important because the monetary-policy framework remains multi-instrument: price-based tools matter more than in the past, but quantity tools, structural facilities, and communication still interact in shaping expectations about financing costs and price recovery. For an economy integrated into global trade and finance, the relevant expectation state is therefore not purely domestic. It is formed at the intersection of domestic demand conditions, external pricing pressures, and policy communication about how the central bank interprets both. Standard empirical tools are not fully adequate for this task. Survey measures are expensive and low frequency. Search data are timely but noisy. Conventional text-as-data approaches often collapse everything into broad positive or negative sentiment. That is too coarse for a policy-facing inflation application, especially in an open economy. A paragraph about pork prices, a report on housing weakness, an article on U.S.–China tariff risk, and a PBOC statement about lowering funding costs all contain price language, but they operate through different economic mechanisms. What matters for inflation monitoring is not generic tone; it is the mechanism through which a narrative affects price expectations, whether through domestic demand, upstream cost pass-through, external prices, exchange-rate channels, or policy credibility. This paper addresses that problem with an explainable framework that treats narratives, attention, and macro variables as complementary inputs. In the full production design, contextual LLM classification maps texts into economically meaningful inflation mechanisms, including demand weakness, imported inflation, exchange-rate pass-through, industrial-price transmission, property spillovers, and policy-support language. Search data measure salience directly, and mixed-frequency methods combine those signals with conventional macro indicators. The empirical implementation in this article is intentionally more conservative: it evaluates a reproducible public-data pilot that uses official Chinese macro series together with transparent communication and attention proxies to test whether an expectation layer improves current-month CPI nowcasting. The empirical contribution is therefore disciplined but non-trivial. Using monthly data for 2023M1–2025M12, the pilot evaluates recursive current-month CPI nowcasts in a low-inflation regime. Models that add communication and a lagged attention proxy outperform stricter macro-only linear benchmarks. Because the sample is short and the text and attention branches are represented by reduced-form proxies, the evidence should be read as a proof-of-concept for policy-relevant measurement rather than as a definitive account of China’s inflation process. Even so, the exercise is informative for open-economy macroeconomics because it shows how domestic and external narratives can be organized as measurable expectation channels, then disciplined by official macro data and evaluated with transparent predictive metrics. The paper makes four contributions. First, it defines an LLM-ready framework that links narrative measurement, public attention, and macroeconomic modelling in a way that is usable for open-economy inflation analysis. Second, it proposes a China-specific inflation-narrative taxonomy organized by economic mechanism rather than generic sentiment, with particular emphasis on imported inflation, exchange-rate narratives, trade-policy news, and policy communication. Third, it provides a reproducible public-data pilot that preserves real-time information discipline and can be replicated with official series. Fourth, it evaluates the framework with explainability tools and a cautious policy-window validation exercise, thereby keeping the empirical contribution proportionate to the available data. The rest of the paper proceeds as follows. Section 2 reviews the literature and states the paper’s contribution. Section 3 explains why China is an appropriate institutional laboratory. Section 4 presents the framework and the narrative taxonomy. Section 5 describes the reproducible public-data pilot. Section 6 outlines the empirical design. Section 7 reports the empirical results. Section 8 discusses robustness, limitations, and deployment. Section 9 concludes. 2 Related Literature and Contribution The paper connects five strands of research that matter for a Journal of International Economics audience: inflation-expectation formation, text and LLM applications in economics, China inflation forecasting, central-bank communication, and explainable AI. Its novelty lies not in importing one isolated technique from each strand, but in combining them in a coherent, policy-relevant workflow for a large open economy in which domestic and external narratives jointly shape inflation expectations. First, a large literature shows that inflation expectations are behaviorally consequential and central to monetary transmission. D’Acunto et al. (2024) survey recent evidence and emphasize that household expectations are heterogeneous, often imperfectly informed, and relevant for policy even when expert or market-based measures are available. In the Chinese context, Li, Sun, and Qiu (2025) show that inflation expectations are associated with heterogeneous consumption intentions across income and age groups. These studies establish why expectations matter, but not how they should be measured in real time from unstructured information flows or how domestic and external narratives should be distinguished when monitoring inflation in an open economy. Second, LLM-based measurement has recently opened a promising route for that task. Jaworski (2025) shows that generative AI can distinguish forward-looking inflation expectations from backward-looking inflation perceptions in online discourse. Zarifhonarvar (2026) extends the frontier by using LLMs in survey-style experiments to generate heterogeneous inflation expectations under alternative information environments. More broadly, recent Computational Economics contributions show growing interest in combining large language models, economic forecasting, and explainability, including in GDP forecasting and interpretable hybrid systems (Darwish et al., 2025; Yao, 2025; Yin & Guo, 2026; Kang et al., 2026). For the present paper, the key lesson is not that LLMs should replace economic structure. It is that they can classify narrative states that matter precisely because open-economy inflation is communicated through rich text about commodities, exchange rates, trade policy, and domestic policy responses. Third, recent work on inflation nowcasting emphasizes the need to combine high-frequency information with disciplined mixed-frequency methods. Knotek and Zaman (2024) survey the nowcasting problem and show why timely information matters before official inflation releases are complete. Schnorrenberger, Schmidt, and Moura (2024) show that machine-learning methods can improve real-time inflation nowcasts when release timing and variable selection are handled carefully. For China specifically, Huang, Qi, and Xia (2024) demonstrate that data-rich machine-learning methods can improve inflation forecasting relative to smaller benchmark models. What remains less developed is a framework in which the expectation layer itself is measured from narratives and attention, then embedded in an interpretable nowcasting system. Fourth, communication is an economically meaningful channel rather than a cosmetic afterthought. Carotta, Mello Costa, and Ponce (2023) show that tone and readability shape firms’ inflation expectations. In the China setting, Su et al. (2025) extract policy signals from PBOC minutes and report summaries, while Wang and Wang (2024) show that PBOC communication sentiment affects long-run stock-bond correlations. These papers indicate that Chinese central-bank communication is measurable and market-relevant, but they stop short of embedding it in a broader inflation-expectation monitoring architecture that also accommodates exchange-rate and imported-inflation narratives. Fifth, explainability has become increasingly important in policy-sensitive machine-learning applications. Monje, Carrasco, and Sánchez-Montañés (2025), Yao (2025), and Kang et al. (2026) all illustrate, in different contexts, that predictive performance without interpretability is unlikely to be persuasive in regulated or policy-facing domains. For inflation monitoring, the question is not only whether a model forecasts well, but also whether it can explain whether the revision came from food-price news, upstream disinflation, imported-cost pressure, housing stress, or changes in policy communication. Taken together, these strands motivate a China-specific applied contribution to international macroeconomics. The framework measures narratives by economic mechanism rather than generic sentiment, embeds the resulting expectation layer in an interpretable nowcasting system, and evaluates it under conservative public-data constraints. The claim is intentionally limited but important: even before a full historical LLM archive is assembled, a carefully constructed expectation layer can add information to inflation monitoring in a large open economy where domestic conditions, external prices, and policy communication interact. 3 Why China Is the Right Institutional Laboratory A China-centered framework is not simply a local variant of a U.S.-style inflation paper. It changes the underlying policy question. In the United States and parts of Europe after the pandemic, the central concern was whether expectations would become unanchored upward. In China during 2023–2025, the immediate concern was whether weak demand, a property-market correction, and persistent PPI weakness would keep prices too soft. That difference matters because the semantics of public narratives are regime dependent. In a high-inflation economy, more price attention often signals overheating fears. In China’s recent regime, the same increase in attention can instead reflect concern about insufficient demand, weak confidence, or easier policy ahead. It can also reflect worries about imported deflationary pressure, global commodity softness, or exchange-rate and tariff developments that alter the path from external prices to domestic inflation. China’s monetary-policy framework also makes communication unusually informative. Maher (2024) shows that although price-based instruments have become more important, the PBOC continues to operate through a broader and more heterogeneous toolkit than a canonical inflation-targeting central bank. The seven-day reverse repo rate, the MLF, the reserve requirement ratio, structural tools, and administrative guidance all matter. In such a framework, official reports, meeting statements, and speeches convey information not fully summarized by one policy rate. That feature is especially important in an open economy, where the same domestic inflation print may have different implications depending on whether policymakers are more concerned about weak domestic demand, imported inflation, exchange-rate stability, or external demand. The official documents themselves underscore that point. The PBOC’s Q4 2024 Monetary Policy Report emphasized an appropriately accommodative stance and linked monetary operations to the goal of supporting a reasonable recovery in prices. The Q4 2024 Monetary Policy Committee statement likewise stressed insufficient domestic demand, potential risks, and the need to improve financing-cost transmission. These documents are not rhetorical decorations; they contain information about the central bank’s diagnosis of the inflation-growth trade-off and about the environment in which households, firms, and markets form expectations. In an internationally integrated economy, that environment also includes how external prices, trade frictions, and exchange-rate movements are narrated and interpreted. China is also a useful methodological case because a public, long, harmonized panel of real-time expectation measures is harder to assemble than in some advanced economies. That constraint strengthens rather than weakens the case for alternative measurement. If an interpretable narrative-based system can recover useful signals under Chinese data conditions, the broader methodology is likely to matter for other middle-income economies facing similar interactions between domestic weakness, external pricing conditions, and limited public expectation data. 4 Framework: From Narrative Classification to Policy-Relevant Nowcasting 4.1 Production design and the role of an LLM-ready architecture The core design principle is straightforward: inflation expectations are formed in language, expressed in attention, disciplined by macro conditions, and revealed in policy and markets. A useful real-time system should therefore use all four dimensions. Figure 1 summarizes the proposed architecture. The text branch ingests official communication, policy reports, news items, and other economically relevant texts. The attention branch measures which topics have become salient in public information demand. The macro branch provides nominal discipline through price, money, credit, and policy variables. The explainability branch converts model output into mechanism-level narratives and counterfactuals. For a country such as China, the architecture is explicitly meant to accommodate both domestic narratives and open-economy narratives, including imported inflation, exchange-rate pass-through, and trade-policy news. Crucially, the framework distinguishes between the production design and the empirical pilot. In production, contextual LLM classification would assign each incoming text unit to economically meaningful narrative classes. Query-level search series would be used directly. Ragged-edge mixed-frequency methods would respect the release calendar. In the pilot, those modules are approximated by transparent public-data stand-ins. That separation is not a weakness; it is what allows the paper to stay reproducible while still proposing a forward-looking research architecture that can later be extended to richer text corpora, search data, and international-price information. 4.2 Narrative taxonomy for a China deployment The text branch is designed to classify documents by mechanism, not by generic emotional polarity. In the China context, a useful taxonomy should at minimum distinguish domestic-demand weakness, food-price shocks, industrial-price pass-through, imported inflation, property spillovers, labor-cost pressure, policy-support language, nominal-anchor language, and explicit deflation concern. Each document may load on more than one class. A PBOC report paragraph can simultaneously communicate weak demand, a desire to lower financing costs, and concern about expectations. A news item can simultaneously discuss commodity prices, exchange-rate pass-through, and tariff pressure. Table 1 presents the working taxonomy used to organize the framework. Table 1. Illustrative inflation-narrative taxonomy for a China deployment. Narrative label Illustrative meaning Policy relevance Demand weakness / consumption restraint Insufficient domestic demand, cautious spending, weak private confidence Signals downside inflation risk and possible support for easing Producer-price deflation / upstream pressure Factory-gate weakness, excess capacity, margin compression Maps upstream disinflation into downstream CPI risk Imported inflation / external cost channel Exchange-rate pass-through, tariffs, or commodity-price pressure Captures externally generated inflation pressure Food and energy relative-price shock Pork, grain, vegetable, fuel, or electricity price news Separates temporary headline shocks from broad inflation pressure Property-market spillover Housing adjustment, wealth effects, mortgage conditions Links real-estate weakness to demand and price persistence Labor-cost / wage pressure Wages, labor shortages, service-cost pressure Captures domestically generated service inflation Policy-support / easing narrative Credit support, lower funding costs, demand stabilization Measures communication linked to reflation support Nominal-anchor / anti-inflation credibility Price stability, anti-inflation commitment, credibility language Measures anchoring force in communication Deflation concern / lowflation persistence Explicit discussion of low inflation or entrenched weakness Critical for China’s recent regime interpretation 4.3 Fusion and explainability The general nowcasting object can be written as: πₜ = F(Eₜ[text], Aₜ, Mₜ; Θₜ). where Eₜ[text] is the text-based expectation layer, Aₜ is the attention layer, and Mₜ is the macro information set. The explainability requirement is not optional. The system must tell users whether the nowcast moved because of food-price news, upstream disinflation, imported-cost pressure, a shift in policy communication, or some interaction among these channels. In the pilot, explainability is implemented through benchmark comparison, permutation-based feature importance, and local scenario perturbations. In a full deployment, those diagnostics could be supplemented by topic-level decomposition and local explanation tools. The point is not to turn international macro monitoring into a black box, but to recover a transparent mapping from narratives to inflation assessments. 5 Data, Variable Construction, and the Reproducible China Pilot 5.1 Motivation and scope of the pilot The empirical pilot asks a narrower question than the full production design: if China inflation is nowcast with official macro data plus transparent communication and attention channels, does the expectation layer add measurable value? The sample runs from 2023M1 to 2025M12. This is a short window, but it covers the post-reopening transition, several periods of CPI weakness, persistent negative PPI inflation, multiple LPR adjustments, and a policy environment increasingly focused on supporting demand and price recovery. The pilot is not a reduced-form substitute for the full international narrative architecture. It is a transparent empirical test of whether expectation-sensitive information can add value even before richer text and search data are assembled. The point of using a simplified pilot is not to pretend that proxies are perfect substitutes for an actual historical LLM archive or archived search-query data. It is to test whether the underlying architecture contains a real signal under conservative conditions. If a transparent public-data version already improves monitoring in a difficult low-inflation environment, then the case for a richer live system becomes stronger. Conversely, if the reduced-form pilot added nothing, there would be little reason to believe that a more elaborate narrative-measurement system would matter. The pilot therefore serves as an empirical discipline device for the broader framework. Table 2. Variables used in the reproducible China public-data pilot. Variable Frequency Role / construction Public source CPI Monthly Headline consumer price index, same month previous year = 100 NBS PPI Monthly Industrial producer price index, year-on-year NBS M2 growth Monthly Broad money growth, year-on-year PBOC 1Y LPR Monthly Loan Prime Rate as observable price-based policy signal PBOC Communication score Quarterly → monthly Accommodative policy-communication intensity mapped from official materials PBOC / author coding Lagged attention proxy Monthly Reduced-form proxy for public attention built from lagged CPI, lagged PPI, and lagged LPR-change information Author construction Expectation layer Monthly Combined communication and salience channels used in full models Author construction 5.2 Communication score The communication score is intended to approximate the policy-narrative branch of the production system using only transparent public materials. It is coded quarterly from PBOC monetary-policy reports and Monetary Policy Committee communications, then mapped to months within each quarter. Higher values reflect more explicit language about supporting demand, lowering financing costs, stabilizing expectations, and fostering a reasonable recovery in prices. The score is therefore not a casual sentiment measure; it is an accommodative policy-communication index. In a full deployment, the hand-coded quarterly score would be replaced by paragraph-level LLM classification and source-weighted aggregation. That extension would make it possible to distinguish, within official communication, narratives about domestic demand, exchange-rate pressure, imported inflation, and policy support. 5.3 Lagged attention proxy Because a harmonized open archive of historical search series is difficult to assemble for the whole sample, the pilot uses a transparent lagged attention proxy rather than claiming direct measurement of search attention. The timing discipline is explicit: the proxy is built only from information observable before the forecast origin, principally lagged signs of price weakness, upstream disinflation, and lagged policy-rate changes. Its purpose is not to replace a genuine search panel. Its purpose is to represent, in reproducible reduced form, when inflation weakness and policy support are likely to become more attention-intensive in the public information environment. In a full deployment, this branch would be expanded to query-level measures of inflation, exchange rates, tariffs, commodity prices, employment, and housing. 5.4 Feature design and evaluation window Each main variable enters the model with one- and two-month lags. The feature set is intentionally parsimonious because the sample is short and the empirical objective is to test the incremental role of the expectation layer rather than to maximize performance with a kitchen-sink predictor set. After lag construction, the pilot uses an expanding training window followed by one-step-ahead recursive current-month predictions. The first estimation window contains 18 observations, after which forecasts are generated sequentially. This structure keeps the exercise close to the real-time monitoring problem that motivates the paper. 6 Empirical Design 6.1 Nowcasting setup Let yₜ denote the current-month CPI index. Let Xₜᴹ contain lagged CPI, lagged PPI, lagged M2 growth, and the one-year LPR. Let Xₜᴱ contain the communication score, the lagged attention proxy, and their lags. The forecasting object is: yₜ = f(Xₜᴹ, Xₜᴱ) + εₜ. The benchmark hierarchy is intentionally strict. The first model is an AR(1) using only lagged CPI. The second and third models are macro-only linear and macro-only elastic-net specifications. The remaining models add the expectation layer: a full elastic net, a gradient-boosting regressor, and a random forest. The parsimonious design is deliberate. In a short sample, a variable that improves performance despite severe dimensional restraint is more informative than one that only helps in a high-dimensional kitchen-sink model. The exercise is therefore not framed as a pure machine-learning horse race; it is a test of whether narrative-sensitive measurement adds disciplined predictive content. 6.2 Why compare linear and tree-based learners? Linear models provide an interpretable benchmark and reveal whether communication and the lagged attention proxy still matter after conventional macro controls are introduced. Tree-based models are included because the China setting is likely to be state dependent. The same degree of inflation attention can mean very different things depending on whether it appears alongside food-price shocks, upstream deflation, housing weakness, or supportive policy language. A flexible learner can capture such interactions, but its output must still be explainable. That trade-off between flexibility and interpretability is central to the design. 6.3 Exploratory policy-window validation The paper also considers whether the same information set lines up with easing windows. Let D_t be an indicator for months in which the one-year LPR is cut. The reduced-form validation equation is: Pr(Dₜ = 1) = Λ(α + β₁CPIₜ₋₁ + β₂PPIₜ₋₁ + β₃M2ₜ₋₁ + β₄COMMₜ₋₁ + β₅ATTNₜ₋₁). This module is explicitly exploratory. With only five observed easing episodes, it should not be interpreted as a structural estimate of the PBOC reaction function. Its value lies in checking whether the same narrative and attention environment that helps nowcast inflation is also aligned with policy-relevant periods. In other words, the exercise asks whether the expectation layer is merely statistically useful or whether it also corresponds to the informational environment in which policy adjustments occur. 6.4 Explainability tools Explainability is built into the design in three ways. First, linear models remain in the benchmark set. Second, for the best-performing nonlinear model the paper reports permutation-based global feature importance. Third, the paper presents local scenario perturbations that translate model behavior into policy language. Those perturbations are illustrative local changes in model output, not structural policy elasticities. Together, these tools make the framework usable for real-time monitoring rather than only for retrospective fit. 7 Empirical Results 7.1 Descriptive dynamics The descriptive evidence already motivates the framework. CPI remains close to the 100 threshold for much of the sample, while PPI stays negative for an extended period. The monetary environment also exhibits a recognizable easing pattern, with the one-year LPR declining in discrete steps and money growth moving from low double digits toward the upper-single-digit range before stabilizing. This combination is consistent with a regime in which inflation expectations are shaped less by overheating fears than by the interaction of weak domestic pricing power, policy support, and the way external and domestic signals are interpreted. The communication score trends upward through the sample, consistent with a policy environment increasingly focused on weak demand, financing conditions, and price recovery. The lagged attention proxy is more volatile, which is exactly what one would expect from a public-attention channel. Communication captures the official policy narrative; the lagged attention proxy captures when inflation weakness and related policy concerns become information-intensive. The two are related but not identical, which is precisely why the framework treats them as complementary rather than interchangeable expectation channels. 7.2 Pseudo-real-time current-month nowcasting performance Table 3 reports the recursive performance comparison. The best RMSE is achieved by the full random forest, followed closely by the full gradient-boosting model. The AR(1) benchmark remains strong, which is unsurprising in a smooth inflation regime, but the full models still improve on the macro-only linear specifications. The right interpretation is not that the pilot has discovered a definitive forecasting champion. It is that, even under severe data and sample constraints, the expectation layer appears to contain information not fully captured by simple macro persistence. Table 3. Recursive current-month nowcasting performance in the China pilot. Model MAE RMSE Test observations Full random forest 0.181 0.253 16 Full gradient boosting 0.180 0.257 16 AR(1) 0.187 0.264 16 Full elastic net 0.213 0.270 16 Macro elastic net 0.218 0.273 16 Macro linear 0.308 0.359 16 Relative to the macro-only linear benchmark, the best full random-forest model lowers RMSE materially. Relative to the macro-only elastic-net benchmark, the gain is smaller but still meaningful. These magnitudes should not be oversold in a short sample, yet they are economically relevant because the pilot stacks the deck against the new variables: the target is smooth, the sample is small, and the narrative channels are measured conservatively. The result therefore supports the paper’s central claim that expectation-sensitive measurement can improve monitoring even before richer open-economy narrative data are introduced. 7.3 Feature attribution The global attribution results are deliberately reassuring. Lagged CPI, lagged PPI, and lagged M2 growth remain the leading features, which is exactly how a credible inflation nowcast should behave. At the same time, the communication and attention variables are not negligible. They contribute meaningfully to the best-performing nonlinear model, which suggests that the expectation layer refines rather than overwhelms the hard macro anchor. Table 4. Global feature contributions in the best-performing nonlinear nowcast. Feature Mean absolute contribution cpi_lag1 0.181 ppi_lag1 0.160 m2_lag1 0.102 m2_lag2 0.078 ppi_lag2 0.075 attention_lag1 0.060 cpi_lag2 0.051 attention_lag2 0.034 comm_lag1 0.030 comm_lag2 0.009 This ranking also suggests a plausible division of labor between the two expectation channels. The lagged attention proxy appears more important than communication in short-run prediction, which is consistent with the idea that public attention spikes are episodic. The communication index is smoother and likely operates more as a policy backdrop than as a month-to-month shock. That pattern fits the institutional setting: official communication evolves gradually, while public attention can jump rapidly when inflation weakness, external pricing news, or policy expectations become salient. 7.4 Local scenario perturbations Table 5 reports local perturbations around the final nowcast. These should be read as model-based scenario diagnostics, not as structural policy multipliers. Two results are especially noteworthy. First, a less deflationary upstream price signal lifts the nowcast, which is economically intuitive. Second, higher attention slightly lowers the nowcast in this sample. That sign is not paradoxical once the Chinese regime is understood: more inflation-related attention in 2023–2025 often reflects concern about weak demand, disinflation, or policy support, rather than fear of overheating. Table 5. Illustrative local scenario perturbations for the final nowcast. Scenario Predicted CPI index Δ vs. base Higher communication clarity 100.519 -0.006 Higher attention 100.480 -0.044 Higher communication and attention 100.475 -0.050 Less deflationary producer-price signal 100.532 0.007 Stronger money growth 100.520 -0.005 7.5 Exploratory policy-window evidence The policy-validation exercise is intentionally modest. With only 35 monthly observations and five easing events, the logit model cannot identify a structural PBOC reaction function. It is retained only as an auxiliary consistency check asking whether softer price conditions, weaker money growth, and a more supportive communication–attention environment line up descriptively with observed easing windows. Table 6. Exploratory policy-window logit summary and coefficients. Indicator Value AUC (in-sample) 0.733 Accuracy (in-sample) 0.657 Positive easing events 5 Observations 35 Coefficient: ppi_lag1 -0.704 Coefficient: m2_lag1 -0.680 Coefficient: cpi_lag1 -0.395 Coefficient: comm_lag1 -0.381 Coefficient: attention_lag1 -0.274 Table 7. Observed one-year LPR-cut episodes in the China pilot sample. Date CPI PPI M2 LPR Comm. Attn. Expectation index 2023-06 100.0 -5.4 11.3 3.55 0.45 0.752 0.586 2023-08 100.1 -3.0 10.6 3.45 0.60 0.873 0.723 2024-07 100.5 -0.8 6.3 3.35 0.70 0.489 0.605 2024-10 100.3 -2.9 7.5 3.10 0.80 0.615 0.717 2025-05 99.9 -3.3 8.3 3.00 0.85 1.000 0.917 The policy evidence is therefore supportive but limited. It suggests that the expectation layer is aligned with the policy environment, but it does not support structural inference. The primary contribution of the paper remains the nowcasting framework and its explainable validation. That is also the right interpretation for a Journal of International Economics audience: the value of the exercise lies in showing how domestic and external narratives can be organized into a disciplined monitoring architecture, not in claiming that the short pilot fully identifies a policy rule. 8 Robustness, Limitations, and Deployment Roadmap The pilot has clear limits. The sample is short; the communication score is a reduced-form stand-in for paragraph-level historical LLM classification; and the attention measure is a lagged proxy rather than direct query-level search data. The pilot is also monthly and does not yet implement full ragged-edge timing, release surprises, or market-pricing validation. In addition, although the conceptual framework is explicitly open-economy, the reduced-form pilot cannot yet isolate separate empirical effects of exchange-rate narratives, imported inflation, or trade-policy news. These limits define the paper’s scope. The article should be read as a disciplined bridge between conceptual architecture and live deployment. Its contribution lies in showing that narrative-sensitive monitoring can add information even under conservative measurement and short-sample conditions. The proof-of-concept is therefore about measurement design and empirical discipline, not about replacing richer structural or international models. A live China system would proceed in six steps. First, ingest official texts, PBOC reports, meeting statements, major financial-news coverage, and market commentary. Second, classify each text unit with an LLM prompt schema built around the taxonomy in Table 1. Third, collect actual query-level search data, ideally from Baidu Index or comparable Chinese search platforms. Fourth, merge those channels with mixed-frequency macro releases and market prices. Fifth, estimate a fusion model that respects the release calendar and allows narrative variables to enter with state dependence. Sixth, validate the resulting system not only on inflation nowcasts, but also on event windows and market repricing around releases and policy communication. In a Journal of International Economics setting, the natural additional step would be to enrich the international block explicitly with exchange rates, import prices, commodity prices, and external-demand indicators. The most important robustness lesson from the pilot is that the expectation layer should remain multi-channel. Communication and attention are related, but they are not substitutes. Communication reflects the official policy narrative; attention reflects what becomes information-intensive to the public and the market. In some episodes they move together; in others they do not. That distinction would become even more important in a full LLM-and-search deployment, especially once external-price and exchange-rate narratives are measured directly. 9 Conclusion This paper develops an explainable framework for measuring real-time inflation expectations in China and embedding those signals in inflation nowcasting. Its main contribution is to organize narrative measurement, public attention, and conventional macro variables into an interpretable system that is usable for open-economy inflation monitoring. The China setting is particularly revealing because the recent policy challenge has been the management of weak inflation rather than classic overheating. In such a regime, more inflation-related discourse can signal concern about insufficient reflation rather than fear of excessive price growth. In a large open economy, that discourse also includes narratives about exchange rates, tariffs, imported inflation, and external demand. This regime dependence is precisely why generic sentiment scores are inadequate and why a structured narrative taxonomy is needed. The public-data pilot provides initial support for the framework. Expectation-augmented models improve on stricter macro-only linear benchmarks in recursive CPI nowcasting. Feature attribution shows that the new variables refine rather than replace macro fundamentals. The exploratory policy-window exercise offers only descriptive support, but it is consistent with the broader view that narratives and attention capture a policy-relevant dimension of the environment. Future work should extend the framework with historical LLM classification of Chinese text corpora, direct query-level search attention, richer international-price and exchange-rate measures, and a high-frequency market-validation module. Even in reduced form, the present results show why inflation monitoring in China is better treated as a problem of narrative measurement, salience, and explainable data fusion than as a purely mechanical extrapolation of lagged CPI. References Carotta, G., Mello Costa, M., & Ponce, J. (2023). Monetary policy communication and inflation expectations: New evidence about tone and readability. Latin American Journal of Central Banking, 4(3), 100088. https://doi.org/10.1016/j.latcb.2023.100088 Darwish, M., Hassanien, E. E., & Eissa, A. H. B. (2025). Stock market forecasting: From traditional predictive models to large language models. Computational Economics. https://doi.org/10.1007/s10614-025-11024-w D’Acunto, F., Charalambakis, E., Georgarakos, D., Kenny, G., Meyer, J., & Weber, M. (2024). Household inflation expectations: An overview of recent insights for monetary policy. NBER Working Paper No. 32488. https://doi.org/10.3386/w32488 Huang, N., Qi, Y., & Xia, J. (2024). China’s inflation forecasting in a data-rich environment: Based on machine learning algorithms. Applied Economics, 57(17), 1995–2020. https://doi.org/10.1080/00036846.2024.2322572 International Monetary Fund. (2026). IMF Executive Board concludes 2025 Article IV consultation with China. IMF Press Release 26/053. Jaworski, K. (2025). Measuring inflation expectations using artificial intelligence. Computational Economics. https://doi.org/10.1007/s10614-025-11231-5 Kang, Y., Ryu, D., & Webb, R. I. (2026). Uncertainty indicators as key predictors of oil volatility: An interpretable machine learning approach. Computational Economics. https://doi.org/10.1007/s10614-025-11299-z Knotek, E. S., II, & Zaman, S. (2024). Nowcasting inflation. Federal Reserve Bank of Cleveland Working Paper No. 24-06. https://doi.org/10.26509/frbc-wp-202406 Li, H., Sun, J., & Qiu, N. (2025). Do inflation expectations affect consumption intentions? Evidence from a survey of Chinese households. Applied Economics Letters. https://doi.org/10.1080/13504851.2025.2466761 Maher, W. (2024). China’s monetary policy framework and financial market transmission. Reserve Bank of Australia Bulletin, April 2024. Monje, L., Carrasco, R. A., & Sánchez-Montañés, M. (2025). Machine learning XAI for early loan default prediction. Computational Economics. https://doi.org/10.1007/s10614-025-10962-9 National Bureau of Statistics of China. (2026a). Consumer Price Index in December 2025. National Bureau of Statistics of China. National Bureau of Statistics of China. (2026b). Industrial Producer Prices in December 2025. National Bureau of Statistics of China. People’s Bank of China. (2024a). PBOC Monetary Policy Committee holds Q4 2024 meeting. People’s Bank of China. People’s Bank of China. (2025a). China monetary policy report, Q4 2024. People’s Bank of China. Schnorrenberger, R., Schmidt, A., & Moura, G. V. (2024). Harnessing machine learning for real-time inflation nowcasting. De Nederlandsche Bank Working Paper No. 806. Su, S., Ahmad, A. H., Wood, J., & Jia, S. (2025). Monetary policy analysis using natural language processing: Evaluating the People’s Bank of China’s minutes and report summary with the Taylor rule. Economic Modelling, 149, 107121. https://doi.org/10.1016/j.econmod.2025.107121 Wang, Y., & Wang, X. (2024). The role of central bank communication in the long-term stock-bond correlations: Evidence from China. Finance Research Letters, 67, 105893. https://doi.org/10.1016/j.frl.2024.105893 Yao, J. (2025). A fusion method integrated econometrics and deep learning to improve the interpretability of prediction: Evidence from Chinese carbon emissions forecast based on OLS-CNN model. Computational Economics, 66, 2987–3006. https://doi.org/10.1007/s10614-024-10793-0 Yin, M., & Guo, M. (2026). The emotional drive of economic forecasting: GDP predictions based on large language models. Computational Economics. https://doi.org/10.1007/s10614-026-11323-w Zarifhonarvar, A. (2026). Generating inflation expectations with large language models. Journal of Monetary Economics, 157, 103859. https://doi.org/10.1016/j.jmoneco.2025.103859 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9395947","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":621757569,"identity":"10fe5fc0-8f38-4826-b646-fbd3139930cc","order_by":0,"name":"XiaoXi Ma","email":"","orcid":"https://orcid.org/0009-0005-3734-5197","institution":"Tianjin Tianshi College, School of Economics and Management","correspondingAuthor":false,"prefix":"","firstName":"XiaoXi","middleName":"","lastName":"Ma","suffix":""},{"id":621757714,"identity":"e1f5669a-2215-4e0e-aed6-c63e9ebfea47","order_by":1,"name":"Lu Chao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1ElEQVRIie3RMQrCMBTG8RcKyfJqN4lU6BUiQqceJlKIi5tQ3EQKmcS5ongWpUMXcZcuLa4OPYHo2ilxE8x///F4fAAu1w82AFZ1EpO1jtpNY0UooIJurMgO0lxYEohJkZTkBHPN7Qg7qwculEfholfPDKJgeDYQnFVTvCaUko2+H28w2R+kgXCQob9VSD2ia1+DFLWZiNB/lfxzRi9tSTwqsBQUifbsCC5S0aGSlJN8dLxx8y8Bqy7NZ0oZFaztnlkSBaGB9ENqN02PfCtcLpfrH3oDas07LIPJjM0AAAAASUVORK5CYII=","orcid":"","institution":"Tianjin Tianshi College, School of Economics and Management","correspondingAuthor":true,"prefix":"","firstName":"Lu","middleName":"","lastName":"Chao","suffix":""},{"id":621757950,"identity":"40221b40-8846-4a0b-ada3-8636f202f8ca","order_by":2,"name":"Shulun Hou","email":"","orcid":"","institution":"School of Artificial Intelligence, Jilin International Studies University","correspondingAuthor":false,"prefix":"","firstName":"Shulun","middleName":"","lastName":"Hou","suffix":""},{"id":621757951,"identity":"156e15a6-18ac-45bb-b9e5-09076985f7ca","order_by":3,"name":"Wang MuYao","email":"","orcid":"","institution":"Hainan University, International Business School","correspondingAuthor":false,"prefix":"","firstName":"Wang","middleName":"","lastName":"MuYao","suffix":""},{"id":621757952,"identity":"2fe3d91b-91c3-4a21-923d-87caa6dcfb00","order_by":4,"name":"SaiChen Jing","email":"","orcid":"","institution":"Tianjin Tianshi College, School of Economics and Management","correspondingAuthor":false,"prefix":"","firstName":"SaiChen","middleName":"","lastName":"Jing","suffix":""}],"badges":[],"createdAt":"2026-04-12 17:10:15","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9395947/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9395947/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106871814,"identity":"327b09d2-e644-49e9-84f4-3d497b20ac3f","added_by":"auto","created_at":"2026-04-14 09:52:31","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":91617,"visible":true,"origin":"","legend":"\u003cp\u003eConceptual architecture of the proposed real-time inflation-expectation system. The submission version implements a transparent reduced-form pilot of this architecture rather than a full historical LLM deployment.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9395947/v1/69afa0f3326af69af58d613b.jpg"},{"id":106994140,"identity":"c8c855ab-1266-4c17-be8e-a52a43b1e984","added_by":"auto","created_at":"2026-04-15 15:05:20","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":77232,"visible":true,"origin":"","legend":"\u003cp\u003eChina pilot data: CPI and PPI dynamics, 2023M1–2025M12.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9395947/v1/7c6a0e2959935b8baef8f1a6.jpg"},{"id":106961086,"identity":"0f329328-7b7e-4a01-9d31-7e6d8e3e214b","added_by":"auto","created_at":"2026-04-15 09:24:13","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":91356,"visible":true,"origin":"","legend":"\u003cp\u003eChina pilot data: M2 growth and the one-year LPR.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9395947/v1/a033d24e53d5a86044a403f0.jpg"},{"id":106961619,"identity":"75d1f439-54bc-4d84-b1e6-29f4be16387f","added_by":"auto","created_at":"2026-04-15 09:26:19","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":97330,"visible":true,"origin":"","legend":"\u003cp\u003eCommunication score and lagged attention proxy.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9395947/v1/a85b74f052d8dd4cbb91af35.jpg"},{"id":106960374,"identity":"3b355b6e-d923-452a-b6de-ea1353203a45","added_by":"auto","created_at":"2026-04-15 09:20:39","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":74913,"visible":true,"origin":"","legend":"\u003cp\u003eRMSE comparison across benchmark and expectation-augmented models\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9395947/v1/53b4ed29a8f1279b46fb15ad.jpg"},{"id":106961934,"identity":"07d199c6-1b78-493d-b987-d4515b52f9a2","added_by":"auto","created_at":"2026-04-15 09:27:52","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":77421,"visible":true,"origin":"","legend":"\u003cp\u003eGlobal explainability ranking for the China pilot nowcast.\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9395947/v1/f98074217e52cf017e33c54b.jpg"},{"id":106961620,"identity":"8e1b4961-eb5e-4acc-a5ca-971149804a98","added_by":"auto","created_at":"2026-04-15 09:26:19","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":96323,"visible":true,"origin":"","legend":"\u003cp\u003eCommunication, attention, and LPR-cut windows in the China pilot.\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9395947/v1/fb7962ef2083092b461dd02c.jpg"},{"id":106994906,"identity":"c61f8a23-23a0-47ee-8038-0eb3bd25a6c9","added_by":"auto","created_at":"2026-04-15 15:20:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1459378,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9395947/v1/d4142185-ded0-412c-b428-9fb3396016ee.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eReal-Time Inflation Expectations in China: An Explainable Nowcasting Framework and a Reproducible Communication–Attention Pilot\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eChina is an especially informative setting for studying real-time inflation expectations in an open economy. During 2023\u0026ndash;2025, the central policy problem was not persistent overheating, but rather whether weak domestic demand, a prolonged property-sector adjustment, and continued producer-price disinflation would keep price recovery too soft for too long. In a large trading economy such as China, that domestic weakness interacts with external cost, exchange-rate, and trade-policy narratives. Imported inflation can arrive through commodities, tariffs, and exchange-rate pass-through, while a weak global pricing environment can reinforce domestic disinflation. The IMF\u0026rsquo;s 2025 Article IV consultation reported that headline inflation averaged 0 percent in 2025 even though GDP growth reached the official target, and the National Bureau of Statistics reported that 2025 CPI was flat on average while producer prices fell 2.6 percent over the year. In that environment, expectations matter because policymakers must prevent a low-inflation equilibrium from becoming self-reinforcing and must judge how domestic and external narratives are affecting price formation in real time.\u003c/p\u003e\n\u003cp\u003eThat policy problem immediately creates a measurement problem. Inflation expectations move faster than official price releases. They respond to food-price headlines, policy-support narratives, property-market news, tariff and exchange-rate stories, and the evolving tone of official communication. In China, these channels are especially important because the monetary-policy framework remains multi-instrument: price-based tools matter more than in the past, but quantity tools, structural facilities, and communication still interact in shaping expectations about financing costs and price recovery. For an economy integrated into global trade and finance, the relevant expectation state is therefore not purely domestic. It is formed at the intersection of domestic demand conditions, external pricing pressures, and policy communication about how the central bank interprets both.\u003c/p\u003e\n\u003cp\u003eStandard empirical tools are not fully adequate for this task. Survey measures are expensive and low frequency. Search data are timely but noisy. Conventional text-as-data approaches often collapse everything into broad positive or negative sentiment. That is too coarse for a policy-facing inflation application, especially in an open economy. A paragraph about pork prices, a report on housing weakness, an article on U.S.\u0026ndash;China tariff risk, and a PBOC statement about lowering funding costs all contain price language, but they operate through different economic mechanisms. What matters for inflation monitoring is not generic tone; it is the mechanism through which a narrative affects price expectations, whether through domestic demand, upstream cost pass-through, external prices, exchange-rate channels, or policy credibility.\u003c/p\u003e\n\u003cp\u003eThis paper addresses that problem with an explainable framework that treats narratives, attention, and macro variables as complementary inputs. In the full production design, contextual LLM classification maps texts into economically meaningful inflation mechanisms, including demand weakness, imported inflation, exchange-rate pass-through, industrial-price transmission, property spillovers, and policy-support language. Search data measure salience directly, and mixed-frequency methods combine those signals with conventional macro indicators. The empirical implementation in this article is intentionally more conservative: it evaluates a reproducible public-data pilot that uses official Chinese macro series together with transparent communication and attention proxies to test whether an expectation layer improves current-month CPI nowcasting.\u003c/p\u003e\n\u003cp\u003eThe empirical contribution is therefore disciplined but non-trivial. Using monthly data for 2023M1\u0026ndash;2025M12, the pilot evaluates recursive current-month CPI nowcasts in a low-inflation regime. Models that add communication and a lagged attention proxy outperform stricter macro-only linear benchmarks. Because the sample is short and the text and attention branches are represented by reduced-form proxies, the evidence should be read as a proof-of-concept for policy-relevant measurement rather than as a definitive account of China\u0026rsquo;s inflation process. Even so, the exercise is informative for open-economy macroeconomics because it shows how domestic and external narratives can be organized as measurable expectation channels, then disciplined by official macro data and evaluated with transparent predictive metrics.\u003c/p\u003e\n\u003cp\u003eThe paper makes four contributions. First, it defines an LLM-ready framework that links narrative measurement, public attention, and macroeconomic modelling in a way that is usable for open-economy inflation analysis. Second, it proposes a China-specific inflation-narrative taxonomy organized by economic mechanism rather than generic sentiment, with particular emphasis on imported inflation, exchange-rate narratives, trade-policy news, and policy communication. Third, it provides a reproducible public-data pilot that preserves real-time information discipline and can be replicated with official series. Fourth, it evaluates the framework with explainability tools and a cautious policy-window validation exercise, thereby keeping the empirical contribution proportionate to the available data.\u003c/p\u003e\n\u003cp\u003eThe rest of the paper proceeds as follows. Section 2 reviews the literature and states the paper\u0026rsquo;s contribution. Section 3 explains why China is an appropriate institutional laboratory. Section 4 presents the framework and the narrative taxonomy. Section 5 describes the reproducible public-data pilot. Section 6 outlines the empirical design. Section 7 reports the empirical results. Section 8 discusses robustness, limitations, and deployment. Section 9 concludes.\u003c/p\u003e"},{"header":"2 Related Literature and Contribution","content":"\u003cp\u003eThe paper connects five strands of research that matter for a Journal of International Economics audience: inflation-expectation formation, text and LLM applications in economics, China inflation forecasting, central-bank communication, and explainable AI. Its novelty lies not in importing one isolated technique from each strand, but in combining them in a coherent, policy-relevant workflow for a large open economy in which domestic and external narratives jointly shape inflation expectations.\u003c/p\u003e\n\u003cp\u003eFirst, a large literature shows that inflation expectations are behaviorally consequential and central to monetary transmission. D\u0026rsquo;Acunto et al. (2024) survey recent evidence and emphasize that household expectations are heterogeneous, often imperfectly informed, and relevant for policy even when expert or market-based measures are available. In the Chinese context, Li, Sun, and Qiu (2025) show that inflation expectations are associated with heterogeneous consumption intentions across income and age groups. These studies establish why expectations matter, but not how they should be measured in real time from unstructured information flows or how domestic and external narratives should be distinguished when monitoring inflation in an open economy.\u003c/p\u003e\n\u003cp\u003eSecond, LLM-based measurement has recently opened a promising route for that task. Jaworski (2025) shows that generative AI can distinguish forward-looking inflation expectations from backward-looking inflation perceptions in online discourse. Zarifhonarvar (2026) extends the frontier by using LLMs in survey-style experiments to generate heterogeneous inflation expectations under alternative information environments. More broadly, recent Computational Economics contributions show growing interest in combining large language models, economic forecasting, and explainability, including in GDP forecasting and interpretable hybrid systems (Darwish et al., 2025; Yao, 2025; Yin \u0026amp; Guo, 2026; Kang et al., 2026). For the present paper, the key lesson is not that LLMs should replace economic structure. It is that they can classify narrative states that matter precisely because open-economy inflation is communicated through rich text about commodities, exchange rates, trade policy, and domestic policy responses.\u003c/p\u003e\n\u003cp\u003eThird, recent work on inflation nowcasting emphasizes the need to combine high-frequency information with disciplined mixed-frequency methods. Knotek and Zaman (2024) survey the nowcasting problem and show why timely information matters before official inflation releases are complete. Schnorrenberger, Schmidt, and Moura (2024) show that machine-learning methods can improve real-time inflation nowcasts when release timing and variable selection are handled carefully. For China specifically, Huang, Qi, and Xia (2024) demonstrate that data-rich machine-learning methods can improve inflation forecasting relative to smaller benchmark models. What remains less developed is a framework in which the expectation layer itself is measured from narratives and attention, then embedded in an interpretable nowcasting system.\u003c/p\u003e\n\u003cp\u003eFourth, communication is an economically meaningful channel rather than a cosmetic afterthought. Carotta, Mello Costa, and Ponce (2023) show that tone and readability shape firms\u0026rsquo; inflation expectations. In the China setting, Su et al. (2025) extract policy signals from PBOC minutes and report summaries, while Wang and Wang (2024) show that PBOC communication sentiment affects long-run stock-bond correlations. These papers indicate that Chinese central-bank communication is measurable and market-relevant, but they stop short of embedding it in a broader inflation-expectation monitoring architecture that also accommodates exchange-rate and imported-inflation narratives.\u003c/p\u003e\n\u003cp\u003eFifth, explainability has become increasingly important in policy-sensitive machine-learning applications. Monje, Carrasco, and S\u0026aacute;nchez-Monta\u0026ntilde;\u0026eacute;s (2025), Yao (2025), and Kang et al. (2026) all illustrate, in different contexts, that predictive performance without interpretability is unlikely to be persuasive in regulated or policy-facing domains. For inflation monitoring, the question is not only whether a model forecasts well, but also whether it can explain whether the revision came from food-price news, upstream disinflation, imported-cost pressure, housing stress, or changes in policy communication.\u003c/p\u003e\n\u003cp\u003eTaken together, these strands motivate a China-specific applied contribution to international macroeconomics. The framework measures narratives by economic mechanism rather than generic sentiment, embeds the resulting expectation layer in an interpretable nowcasting system, and evaluates it under conservative public-data constraints. The claim is intentionally limited but important: even before a full historical LLM archive is assembled, a carefully constructed expectation layer can add information to inflation monitoring in a large open economy where domestic conditions, external prices, and policy communication interact.\u003c/p\u003e"},{"header":"3 Why China Is the Right Institutional Laboratory","content":"\u003cp\u003eA China-centered framework is not simply a local variant of a U.S.-style inflation paper. It changes the underlying policy question. In the United States and parts of Europe after the pandemic, the central concern was whether expectations would become unanchored upward. In China during 2023\u0026ndash;2025, the immediate concern was whether weak demand, a property-market correction, and persistent PPI weakness would keep prices too soft. That difference matters because the semantics of public narratives are regime dependent. In a high-inflation economy, more price attention often signals overheating fears. In China\u0026rsquo;s recent regime, the same increase in attention can instead reflect concern about insufficient demand, weak confidence, or easier policy ahead. It can also reflect worries about imported deflationary pressure, global commodity softness, or exchange-rate and tariff developments that alter the path from external prices to domestic inflation.\u003c/p\u003e\n\u003cp\u003eChina\u0026rsquo;s monetary-policy framework also makes communication unusually informative. Maher (2024) shows that although price-based instruments have become more important, the PBOC continues to operate through a broader and more heterogeneous toolkit than a canonical inflation-targeting central bank. The seven-day reverse repo rate, the MLF, the reserve requirement ratio, structural tools, and administrative guidance all matter. In such a framework, official reports, meeting statements, and speeches convey information not fully summarized by one policy rate. That feature is especially important in an open economy, where the same domestic inflation print may have different implications depending on whether policymakers are more concerned about weak domestic demand, imported inflation, exchange-rate stability, or external demand.\u003c/p\u003e\n\u003cp\u003eThe official documents themselves underscore that point. The PBOC\u0026rsquo;s Q4 2024 Monetary Policy Report emphasized an appropriately accommodative stance and linked monetary operations to the goal of supporting a reasonable recovery in prices. The Q4 2024 Monetary Policy Committee statement likewise stressed insufficient domestic demand, potential risks, and the need to improve financing-cost transmission. These documents are not rhetorical decorations; they contain information about the central bank\u0026rsquo;s diagnosis of the inflation-growth trade-off and about the environment in which households, firms, and markets form expectations. In an internationally integrated economy, that environment also includes how external prices, trade frictions, and exchange-rate movements are narrated and interpreted.\u003c/p\u003e\n\u003cp\u003eChina is also a useful methodological case because a public, long, harmonized panel of real-time expectation measures is harder to assemble than in some advanced economies. That constraint strengthens rather than weakens the case for alternative measurement. If an interpretable narrative-based system can recover useful signals under Chinese data conditions, the broader methodology is likely to matter for other middle-income economies facing similar interactions between domestic weakness, external pricing conditions, and limited public expectation data.\u003c/p\u003e"},{"header":"4 Framework: From Narrative Classification to Policy-Relevant Nowcasting","content":"\u003ch2\u003e4.1 Production design and the role of an LLM-ready architecture\u003c/h2\u003e\n\u003cp\u003eThe core design principle is straightforward: inflation expectations are formed in language, expressed in attention, disciplined by macro conditions, and revealed in policy and markets. A useful real-time system should therefore use all four dimensions. Figure 1 summarizes the proposed architecture. The text branch ingests official communication, policy reports, news items, and other economically relevant texts. The attention branch measures which topics have become salient in public information demand. The macro branch provides nominal discipline through price, money, credit, and policy variables. The explainability branch converts model output into mechanism-level narratives and counterfactuals. For a country such as China, the architecture is explicitly meant to accommodate both domestic narratives and open-economy narratives, including imported inflation, exchange-rate pass-through, and trade-policy news.\u003c/p\u003e\n\u003cp\u003eCrucially, the framework distinguishes between the production design and the empirical pilot. In production, contextual LLM classification would assign each incoming text unit to economically meaningful narrative classes. Query-level search series would be used directly. Ragged-edge mixed-frequency methods would respect the release calendar. In the pilot, those modules are approximated by transparent public-data stand-ins. That separation is not a weakness; it is what allows the paper to stay reproducible while still proposing a forward-looking research architecture that can later be extended to richer text corpora, search data, and international-price information.\u003c/p\u003e\n\u003ch2\u003e4.2 Narrative taxonomy for a China deployment\u003c/h2\u003e\n\u003cp\u003eThe text branch is designed to classify documents by mechanism, not by generic emotional polarity. In the China context, a useful taxonomy should at minimum distinguish domestic-demand weakness, food-price shocks, industrial-price pass-through, imported inflation, property spillovers, labor-cost pressure, policy-support language, nominal-anchor language, and explicit deflation concern. Each document may load on more than one class. A PBOC report paragraph can simultaneously communicate weak demand, a desire to lower financing costs, and concern about expectations. A news item can simultaneously discuss commodity prices, exchange-rate pass-through, and tariff pressure. Table 1 presents the working taxonomy used to organize the framework.\u003c/p\u003e\n\u003cp\u003eTable 1. Illustrative inflation-narrative taxonomy for a China deployment.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eNarrative label\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eIllustrative meaning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePolicy relevance\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eDemand weakness / consumption restraint\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eInsufficient domestic demand, cautious spending, weak private confidence\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSignals downside inflation risk and possible support for easing\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eProducer-price deflation / upstream pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eFactory-gate weakness, excess capacity, margin compression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eMaps upstream disinflation into downstream CPI risk\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eImported inflation / external cost channel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eExchange-rate pass-through, tariffs, or commodity-price pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCaptures externally generated inflation pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eFood and energy relative-price shock\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePork, grain, vegetable, fuel, or electricity price news\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSeparates temporary headline shocks from broad inflation pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eProperty-market spillover\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHousing adjustment, wealth effects, mortgage conditions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eLinks real-estate weakness to demand and price persistence\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eLabor-cost / wage pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eWages, labor shortages, service-cost pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCaptures domestically generated service inflation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePolicy-support / easing narrative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCredit support, lower funding costs, demand stabilization\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eMeasures communication linked to reflation support\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eNominal-anchor / anti-inflation credibility\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePrice stability, anti-inflation commitment, credibility language\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eMeasures anchoring force in communication\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 165px;\"\u003e\n \u003cp\u003eDeflation concern / lowflation persistence\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 237px;\"\u003e\n \u003cp\u003eExplicit discussion of low inflation or entrenched weakness\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 222px;\"\u003e\n \u003cp\u003eCritical for China\u0026rsquo;s recent regime interpretation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003e4.3 Fusion and explainability\u003c/h2\u003e\n\u003cp\u003eThe general nowcasting object can be written as:\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026pi;ₜ = F(Eₜ[text], Aₜ, Mₜ; \u0026Theta;ₜ).\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ewhere Eₜ[text] is the text-based expectation layer, Aₜ is the attention layer, and Mₜ is the macro information set. The explainability requirement is not optional. The system must tell users whether the nowcast moved because of food-price news, upstream disinflation, imported-cost pressure, a shift in policy communication, or some interaction among these channels. In the pilot, explainability is implemented through benchmark comparison, permutation-based feature importance, and local scenario perturbations. In a full deployment, those diagnostics could be supplemented by topic-level decomposition and local explanation tools. The point is not to turn international macro monitoring into a black box, but to recover a transparent mapping from narratives to inflation assessments.\u003c/p\u003e\n\u003ch2\u003e5 Data, Variable Construction, and the Reproducible China Pilot\u003c/h2\u003e\n\u003ch2\u003e5.1 Motivation and scope of the pilot\u003c/h2\u003e\n\u003cp\u003eThe empirical pilot asks a narrower question than the full production design: if China inflation is nowcast with official macro data plus transparent communication and attention channels, does the expectation layer add measurable value? The sample runs from 2023M1 to 2025M12. This is a short window, but it covers the post-reopening transition, several periods of CPI weakness, persistent negative PPI inflation, multiple LPR adjustments, and a policy environment increasingly focused on supporting demand and price recovery. The pilot is not a reduced-form substitute for the full international narrative architecture. It is a transparent empirical test of whether expectation-sensitive information can add value even before richer text and search data are assembled.\u003c/p\u003e\n\u003cp\u003eThe point of using a simplified pilot is not to pretend that proxies are perfect substitutes for an actual historical LLM archive or archived search-query data. It is to test whether the underlying architecture contains a real signal under conservative conditions. If a transparent public-data version already improves monitoring in a difficult low-inflation environment, then the case for a richer live system becomes stronger. Conversely, if the reduced-form pilot added nothing, there would be little reason to believe that a more elaborate narrative-measurement system would matter. The pilot therefore serves as an empirical discipline device for the broader framework.\u003c/p\u003e\n\u003cp\u003eTable 2. Variables used in the reproducible China public-data pilot.\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\" class=\"fr-table-selection-hover\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.2308%;\"\u003e\n \u003cp\u003eVariable\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.0641%;\"\u003e\n \u003cp\u003eFrequency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 306px;\"\u003e\n \u003cp\u003eRole / construction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003ePublic source\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.2308%;\"\u003e\n \u003cp\u003eCPI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.0641%;\"\u003e\n \u003cp\u003eMonthly\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 306px;\"\u003e\n \u003cp\u003eHeadline consumer price index, same month previous year = 100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003eNBS\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.2308%;\"\u003e\n \u003cp\u003ePPI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.0641%;\"\u003e\n \u003cp\u003eMonthly\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 306px;\"\u003e\n \u003cp\u003eIndustrial producer price index, year-on-year\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003eNBS\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.2308%;\"\u003e\n \u003cp\u003eM2 growth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.0641%;\"\u003e\n \u003cp\u003eMonthly\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 306px;\"\u003e\n \u003cp\u003eBroad money growth, year-on-year\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003ePBOC\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.2308%;\"\u003e\n \u003cp\u003e1Y LPR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.0641%;\"\u003e\n \u003cp\u003eMonthly\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 306px;\"\u003e\n \u003cp\u003eLoan Prime Rate as observable price-based policy signal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003ePBOC\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.2308%;\"\u003e\n \u003cp\u003eCommunication score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.0641%;\"\u003e\n \u003cp\u003eQuarterly \u0026rarr; monthly\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 306px;\"\u003e\n \u003cp\u003eAccommodative policy-communication intensity mapped from official materials\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003ePBOC / author coding\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.2308%;\"\u003e\n \u003cp\u003eLagged attention proxy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.0641%;\"\u003e\n \u003cp\u003eMonthly\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 306px;\"\u003e\n \u003cp\u003eReduced-form proxy for public attention built from lagged CPI, lagged PPI, and lagged LPR-change information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003eAuthor construction\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 19.2308%;\"\u003e\n \u003cp\u003eExpectation layer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.0641%;\"\u003e\n \u003cp\u003eMonthly\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 306px;\"\u003e\n \u003cp\u003eCombined communication and salience channels used in full models\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003eAuthor construction\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003ch2\u003e5.2 Communication score\u003c/h2\u003e\n\u003cp\u003eThe communication score is intended to approximate the policy-narrative branch of the production system using only transparent public materials. It is coded quarterly from PBOC monetary-policy reports and Monetary Policy Committee communications, then mapped to months within each quarter. Higher values reflect more explicit language about supporting demand, lowering financing costs, stabilizing expectations, and fostering a reasonable recovery in prices. The score is therefore not a casual sentiment measure; it is an accommodative policy-communication index. In a full deployment, the hand-coded quarterly score would be replaced by paragraph-level LLM classification and source-weighted aggregation. That extension would make it possible to distinguish, within official communication, narratives about domestic demand, exchange-rate pressure, imported inflation, and policy support.\u003c/p\u003e\n\u003ch2\u003e5.3 Lagged attention proxy\u003c/h2\u003e\n\u003cp\u003eBecause a harmonized open archive of historical search series is difficult to assemble for the whole sample, the pilot uses a transparent lagged attention proxy rather than claiming direct measurement of search attention. The timing discipline is explicit: the proxy is built only from information observable before the forecast origin, principally lagged signs of price weakness, upstream disinflation, and lagged policy-rate changes. Its purpose is not to replace a genuine search panel. Its purpose is to represent, in reproducible reduced form, when inflation weakness and policy support are likely to become more attention-intensive in the public information environment. In a full deployment, this branch would be expanded to query-level measures of inflation, exchange rates, tariffs, commodity prices, employment, and housing.\u003c/p\u003e\n\u003ch2\u003e5.4 Feature design and evaluation window\u003c/h2\u003e\n\u003cp\u003eEach main variable enters the model with one- and two-month lags. The feature set is intentionally parsimonious because the sample is short and the empirical objective is to test the incremental role of the expectation layer rather than to maximize performance with a kitchen-sink predictor set. After lag construction, the pilot uses an expanding training window followed by one-step-ahead recursive current-month predictions. The first estimation window contains 18 observations, after which forecasts are generated sequentially. This structure keeps the exercise close to the real-time monitoring problem that motivates the paper.\u003c/p\u003e"},{"header":"6 Empirical Design","content":"\u003ch2\u003e6.1 Nowcasting setup\u003c/h2\u003e\n\u003cp\u003eLet yₜ denote the current-month CPI index. Let Xₜᴹ contain lagged CPI, lagged PPI, lagged M2 growth, and the one-year LPR. Let Xₜᴱ contain the communication score, the lagged attention proxy, and their lags. The forecasting object is:\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eyₜ = f(Xₜᴹ, Xₜᴱ) + \u0026epsilon;ₜ.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe benchmark hierarchy is intentionally strict. The first model is an AR(1) using only lagged CPI. The second and third models are macro-only linear and macro-only elastic-net specifications. The remaining models add the expectation layer: a full elastic net, a gradient-boosting regressor, and a random forest. The parsimonious design is deliberate. In a short sample, a variable that improves performance despite severe dimensional restraint is more informative than one that only helps in a high-dimensional kitchen-sink model. The exercise is therefore not framed as a pure machine-learning horse race; it is a test of whether narrative-sensitive measurement adds disciplined predictive content.\u003c/p\u003e\n\u003ch2\u003e6.2 Why compare linear and tree-based learners?\u003c/h2\u003e\n\u003cp\u003eLinear models provide an interpretable benchmark and reveal whether communication and the lagged attention proxy still matter after conventional macro controls are introduced. Tree-based models are included because the China setting is likely to be state dependent. The same degree of inflation attention can mean very different things depending on whether it appears alongside food-price shocks, upstream deflation, housing weakness, or supportive policy language. A flexible learner can capture such interactions, but its output must still be explainable. That trade-off between flexibility and interpretability is central to the design.\u003c/p\u003e\n\u003ch2\u003e6.3 Exploratory policy-window validation\u003c/h2\u003e\n\u003cp\u003eThe paper also considers whether the same information set lines up with easing windows. Let D_t be an indicator for months in which the one-year LPR is cut. The reduced-form validation equation is:\u003c/p\u003e\n\u003cp\u003e\u003cem\u003ePr(Dₜ = 1) = \u0026Lambda;(\u0026alpha; + \u0026beta;₁CPIₜ₋₁ + \u0026beta;₂PPIₜ₋₁ + \u0026beta;₃M2ₜ₋₁ + \u0026beta;₄COMMₜ₋₁ + \u0026beta;₅ATTNₜ₋₁).\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis module is explicitly exploratory. With only five observed easing episodes, it should not be interpreted as a structural estimate of the PBOC reaction function. Its value lies in checking whether the same narrative and attention environment that helps nowcast inflation is also aligned with policy-relevant periods. In other words, the exercise asks whether the expectation layer is merely statistically useful or whether it also corresponds to the informational environment in which policy adjustments occur.\u003c/p\u003e\n\u003ch2\u003e6.4 Explainability tools\u003c/h2\u003e\n\u003cp\u003eExplainability is built into the design in three ways. First, linear models remain in the benchmark set. Second, for the best-performing nonlinear model the paper reports permutation-based global feature importance. Third, the paper presents local scenario perturbations that translate model behavior into policy language. Those perturbations are illustrative local changes in model output, not structural policy elasticities. Together, these tools make the framework usable for real-time monitoring rather than only for retrospective fit.\u003c/p\u003e"},{"header":"7 Empirical Results","content":"\u003ch2\u003e7.1 Descriptive dynamics\u003c/h2\u003e\n\u003cp\u003eThe descriptive evidence already motivates the framework. CPI remains close to the 100 threshold for much of the sample, while PPI stays negative for an extended period. The monetary environment also exhibits a recognizable easing pattern, with the one-year LPR declining in discrete steps and money growth moving from low double digits toward the upper-single-digit range before stabilizing. This combination is consistent with a regime in which inflation expectations are shaped less by overheating fears than by the interaction of weak domestic pricing power, policy support, and the way external and domestic signals are interpreted.\u003c/p\u003e\n\u003cp\u003eThe communication score trends upward through the sample, consistent with a policy environment increasingly focused on weak demand, financing conditions, and price recovery. The lagged attention proxy is more volatile, which is exactly what one would expect from a public-attention channel. Communication captures the official policy narrative; the lagged attention proxy captures when inflation weakness and related policy concerns become information-intensive. The two are related but not identical, which is precisely why the framework treats them as complementary rather than interchangeable expectation channels.\u003c/p\u003e\n\u003ch2\u003e7.2 Pseudo-real-time current-month nowcasting performance\u003c/h2\u003e\n\u003cp\u003eTable 3 reports the recursive performance comparison. The best RMSE is achieved by the full random forest, followed closely by the full gradient-boosting model. The AR(1) benchmark remains strong, which is unsurprising in a smooth inflation regime, but the full models still improve on the macro-only linear specifications. The right interpretation is not that the pilot has discovered a definitive forecasting champion. It is that, even under severe data and sample constraints, the expectation layer appears to contain information not fully captured by simple macro persistence.\u003c/p\u003e\n\u003cp\u003eTable 3. Recursive current-month nowcasting performance in the China pilot.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 239px;\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003eMAE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003eRMSE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 206px;\"\u003e\n \u003cp\u003eTest observations\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 239px;\"\u003e\n \u003cp\u003eFull random forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.181\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e0.253\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 206px;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 239px;\"\u003e\n \u003cp\u003eFull gradient boosting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.180\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e0.257\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 206px;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 239px;\"\u003e\n \u003cp\u003eAR(1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.187\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e0.264\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 206px;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 239px;\"\u003e\n \u003cp\u003eFull elastic net\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.213\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e0.270\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 206px;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 239px;\"\u003e\n \u003cp\u003eMacro elastic net\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.218\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e0.273\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 206px;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 239px;\"\u003e\n \u003cp\u003eMacro linear\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e0.308\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 96px;\"\u003e\n \u003cp\u003e0.359\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 206px;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eRelative to the macro-only linear benchmark, the best full random-forest model lowers RMSE materially. Relative to the macro-only elastic-net benchmark, the gain is smaller but still meaningful. These magnitudes should not be oversold in a short sample, yet they are economically relevant because the pilot stacks the deck against the new variables: the target is smooth, the sample is small, and the narrative channels are measured conservatively. The result therefore supports the paper\u0026rsquo;s central claim that expectation-sensitive measurement can improve monitoring even before richer open-economy narrative data are introduced.\u003c/p\u003e\n\u003ch2\u003e7.3 Feature attribution\u003c/h2\u003e\n\u003cp\u003eThe global attribution results are deliberately reassuring. Lagged CPI, lagged PPI, and lagged M2 growth remain the leading features, which is exactly how a credible inflation nowcast should behave. At the same time, the communication and attention variables are not negligible. They contribute meaningfully to the best-performing nonlinear model, which suggests that the expectation layer refines rather than overwhelms the hard macro anchor.\u003c/p\u003e\n\u003cp\u003eTable 4. Global feature contributions in the best-performing nonlinear nowcast.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003eFeature\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003eMean absolute contribution\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003ecpi_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.181\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003eppi_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.160\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003em2_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.102\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003em2_lag2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.078\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003eppi_lag2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.075\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003eattention_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.060\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003ecpi_lag2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.051\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003eattention_lag2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.034\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003ecomm_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.030\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003ecomm_lag2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 397px;\"\u003e\n \u003cp\u003e0.009\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThis ranking also suggests a plausible division of labor between the two expectation channels. The lagged attention proxy appears more important than communication in short-run prediction, which is consistent with the idea that public attention spikes are episodic. The communication index is smoother and likely operates more as a policy backdrop than as a month-to-month shock. That pattern fits the institutional setting: official communication evolves gradually, while public attention can jump rapidly when inflation weakness, external pricing news, or policy expectations become salient.\u003c/p\u003e\n\u003ch2\u003e7.4 Local scenario perturbations\u003c/h2\u003e\n\u003cp\u003eTable 5 reports local perturbations around the final nowcast. These should be read as model-based scenario diagnostics, not as structural policy multipliers. Two results are especially noteworthy. First, a less deflationary upstream price signal lifts the nowcast, which is economically intuitive. Second, higher attention slightly lowers the nowcast in this sample. That sign is not paradoxical once the Chinese regime is understood: more inflation-related attention in 2023\u0026ndash;2025 often reflects concern about weak demand, disinflation, or policy support, rather than fear of overheating.\u003c/p\u003e\n\u003cp\u003eTable 5. Illustrative local scenario perturbations for the final nowcast.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 333px;\"\u003e\n \u003cp\u003eScenario\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003ePredicted CPI index\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u0026Delta; vs. base\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 333px;\"\u003e\n \u003cp\u003eHigher communication clarity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e100.519\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e-0.006\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 333px;\"\u003e\n \u003cp\u003eHigher attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e100.480\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e-0.044\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 333px;\"\u003e\n \u003cp\u003eHigher communication and attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e100.475\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e-0.050\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 333px;\"\u003e\n \u003cp\u003eLess deflationary producer-price signal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e100.532\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.007\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 333px;\"\u003e\n \u003cp\u003eStronger money growth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 189px;\"\u003e\n \u003cp\u003e100.520\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e-0.005\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003e7.5 Exploratory policy-window evidence\u003c/h2\u003e\n\u003cp\u003eThe policy-validation exercise is intentionally modest. With only 35 monthly observations and five easing events, the logit model cannot identify a structural PBOC reaction function. It is retained only as an auxiliary consistency check asking whether softer price conditions, weaker money growth, and a more supportive communication\u0026ndash;attention environment line up descriptively with observed easing windows.\u003c/p\u003e\n\u003cp\u003eTable 6. Exploratory policy-window logit summary and coefficients.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003eIndicator\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003eValue\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003eAUC (in-sample)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003e0.733\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003eAccuracy (in-sample)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003e0.657\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003ePositive easing events\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003eObservations\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003e35\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003eCoefficient: ppi_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003e-0.704\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003eCoefficient: m2_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003e-0.680\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003eCoefficient: cpi_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003e-0.395\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003eCoefficient: comm_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003e-0.381\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 486px;\"\u003e\n \u003cp\u003eCoefficient: attention_lag1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 138px;\"\u003e\n \u003cp\u003e-0.274\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable 7. Observed one-year LPR-cut episodes in the China pilot sample.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003eDate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003eCPI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 51px;\"\u003e\n \u003cp\u003ePPI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 54px;\"\u003e\n \u003cp\u003eM2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003eLPR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003eComm.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003eAttn.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003eExpectation index\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e2023-06\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e100.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 51px;\"\u003e\n \u003cp\u003e-5.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 54px;\"\u003e\n \u003cp\u003e11.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003e3.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003e0.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e0.752\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.586\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e2023-08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e100.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 51px;\"\u003e\n \u003cp\u003e-3.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 54px;\"\u003e\n \u003cp\u003e10.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003e3.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e0.873\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.723\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e2024-07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e100.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 51px;\"\u003e\n \u003cp\u003e-0.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 54px;\"\u003e\n \u003cp\u003e6.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003e3.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e0.489\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.605\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e2024-10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e100.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 51px;\"\u003e\n \u003cp\u003e-2.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 54px;\"\u003e\n \u003cp\u003e7.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003e3.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e0.615\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.717\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 83px;\"\u003e\n \u003cp\u003e2025-05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e99.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 51px;\"\u003e\n \u003cp\u003e-3.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 54px;\"\u003e\n \u003cp\u003e8.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003e3.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 63px;\"\u003e\n \u003cp\u003e1.000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 171px;\"\u003e\n \u003cp\u003e0.917\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe policy evidence is therefore supportive but limited. It suggests that the expectation layer is aligned with the policy environment, but it does not support structural inference. The primary contribution of the paper remains the nowcasting framework and its explainable validation. That is also the right interpretation for a Journal of International Economics audience: the value of the exercise lies in showing how domestic and external narratives can be organized into a disciplined monitoring architecture, not in claiming that the short pilot fully identifies a policy rule.\u003c/p\u003e"},{"header":"8 Robustness, Limitations, and Deployment Roadmap","content":"\u003cp\u003eThe pilot has clear limits. The sample is short; the communication score is a reduced-form stand-in for paragraph-level historical LLM classification; and the attention measure is a lagged proxy rather than direct query-level search data. The pilot is also monthly and does not yet implement full ragged-edge timing, release surprises, or market-pricing validation. In addition, although the conceptual framework is explicitly open-economy, the reduced-form pilot cannot yet isolate separate empirical effects of exchange-rate narratives, imported inflation, or trade-policy news.\u003c/p\u003e\n\u003cp\u003eThese limits define the paper\u0026rsquo;s scope. The article should be read as a disciplined bridge between conceptual architecture and live deployment. Its contribution lies in showing that narrative-sensitive monitoring can add information even under conservative measurement and short-sample conditions. The proof-of-concept is therefore about measurement design and empirical discipline, not about replacing richer structural or international models.\u003c/p\u003e\n\u003cp\u003eA live China system would proceed in six steps. First, ingest official texts, PBOC reports, meeting statements, major financial-news coverage, and market commentary. Second, classify each text unit with an LLM prompt schema built around the taxonomy in Table 1. Third, collect actual query-level search data, ideally from Baidu Index or comparable Chinese search platforms. Fourth, merge those channels with mixed-frequency macro releases and market prices. Fifth, estimate a fusion model that respects the release calendar and allows narrative variables to enter with state dependence. Sixth, validate the resulting system not only on inflation nowcasts, but also on event windows and market repricing around releases and policy communication. In a Journal of International Economics setting, the natural additional step would be to enrich the international block explicitly with exchange rates, import prices, commodity prices, and external-demand indicators.\u003c/p\u003e\n\u003cp\u003eThe most important robustness lesson from the pilot is that the expectation layer should remain multi-channel. Communication and attention are related, but they are not substitutes. Communication reflects the official policy narrative; attention reflects what becomes information-intensive to the public and the market. In some episodes they move together; in others they do not. That distinction would become even more important in a full LLM-and-search deployment, especially once external-price and exchange-rate narratives are measured directly.\u003c/p\u003e"},{"header":"9 Conclusion","content":"\u003cp\u003eThis paper develops an explainable framework for measuring real-time inflation expectations in China and embedding those signals in inflation nowcasting. Its main contribution is to organize narrative measurement, public attention, and conventional macro variables into an interpretable system that is usable for open-economy inflation monitoring.\u003c/p\u003e\n\u003cp\u003eThe China setting is particularly revealing because the recent policy challenge has been the management of weak inflation rather than classic overheating. In such a regime, more inflation-related discourse can signal concern about insufficient reflation rather than fear of excessive price growth. In a large open economy, that discourse also includes narratives about exchange rates, tariffs, imported inflation, and external demand. This regime dependence is precisely why generic sentiment scores are inadequate and why a structured narrative taxonomy is needed.\u003c/p\u003e\n\u003cp\u003eThe public-data pilot provides initial support for the framework. Expectation-augmented models improve on stricter macro-only linear benchmarks in recursive CPI nowcasting. Feature attribution shows that the new variables refine rather than replace macro fundamentals. The exploratory policy-window exercise offers only descriptive support, but it is consistent with the broader view that narratives and attention capture a policy-relevant dimension of the environment.\u003c/p\u003e\n\u003cp\u003eFuture work should extend the framework with historical LLM classification of Chinese text corpora, direct query-level search attention, richer international-price and exchange-rate measures, and a high-frequency market-validation module. Even in reduced form, the present results show why inflation monitoring in China is better treated as a problem of narrative measurement, salience, and explainable data fusion than as a purely mechanical extrapolation of lagged CPI.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCarotta, G., Mello Costa, M., \u0026amp; Ponce, J. (2023). Monetary policy communication and inflation expectations: New evidence about tone and readability. Latin American Journal of Central Banking, 4(3), 100088. https://doi.org/10.1016/j.latcb.2023.100088\u003c/li\u003e\n\u003cli\u003eDarwish, M., Hassanien, E. E., \u0026amp; Eissa, A. H. B. (2025). Stock market forecasting: From traditional predictive models to large language models. Computational Economics. https://doi.org/10.1007/s10614-025-11024-w\u003c/li\u003e\n\u003cli\u003eD\u0026rsquo;Acunto, F., Charalambakis, E., Georgarakos, D., Kenny, G., Meyer, J., \u0026amp; Weber, M. (2024). Household inflation expectations: An overview of recent insights for monetary policy. NBER Working Paper No. 32488. https://doi.org/10.3386/w32488\u003c/li\u003e\n\u003cli\u003eHuang, N., Qi, Y., \u0026amp; Xia, J. (2024). China\u0026rsquo;s inflation forecasting in a data-rich environment: Based on machine learning algorithms. Applied Economics, 57(17), 1995\u0026ndash;2020. https://doi.org/10.1080/00036846.2024.2322572\u003c/li\u003e\n\u003cli\u003eInternational Monetary Fund. (2026). IMF Executive Board concludes 2025 Article IV consultation with China. IMF Press Release 26/053.\u003c/li\u003e\n\u003cli\u003eJaworski, K. (2025). Measuring inflation expectations using artificial intelligence. Computational Economics. https://doi.org/10.1007/s10614-025-11231-5\u003c/li\u003e\n\u003cli\u003eKang, Y., Ryu, D., \u0026amp; Webb, R. I. (2026). Uncertainty indicators as key predictors of oil volatility: An interpretable machine learning approach. Computational Economics. https://doi.org/10.1007/s10614-025-11299-z\u003c/li\u003e\n\u003cli\u003eKnotek, E. S., II, \u0026amp; Zaman, S. (2024). Nowcasting inflation. Federal Reserve Bank of Cleveland Working Paper No. 24-06. https://doi.org/10.26509/frbc-wp-202406\u003c/li\u003e\n\u003cli\u003eLi, H., Sun, J., \u0026amp; Qiu, N. (2025). Do inflation expectations affect consumption intentions? Evidence from a survey of Chinese households. Applied Economics Letters. https://doi.org/10.1080/13504851.2025.2466761\u003c/li\u003e\n\u003cli\u003eMaher, W. (2024). China\u0026rsquo;s monetary policy framework and financial market transmission. Reserve Bank of Australia Bulletin, April 2024.\u003c/li\u003e\n\u003cli\u003eMonje, L., Carrasco, R. A., \u0026amp; S\u0026aacute;nchez-Monta\u0026ntilde;\u0026eacute;s, M. (2025). Machine learning XAI for early loan default prediction. Computational Economics. https://doi.org/10.1007/s10614-025-10962-9\u003c/li\u003e\n\u003cli\u003eNational Bureau of Statistics of China. (2026a). Consumer Price Index in December 2025. National Bureau of Statistics of China.\u003c/li\u003e\n\u003cli\u003eNational Bureau of Statistics of China. (2026b). Industrial Producer Prices in December 2025. National Bureau of Statistics of China.\u003c/li\u003e\n\u003cli\u003ePeople\u0026rsquo;s Bank of China. (2024a). PBOC Monetary Policy Committee holds Q4 2024 meeting. People\u0026rsquo;s Bank of China.\u003c/li\u003e\n\u003cli\u003ePeople\u0026rsquo;s Bank of China. (2025a). China monetary policy report, Q4 2024. People\u0026rsquo;s Bank of China.\u003c/li\u003e\n\u003cli\u003eSchnorrenberger, R., Schmidt, A., \u0026amp; Moura, G. V. (2024). Harnessing machine learning for real-time inflation nowcasting. De Nederlandsche Bank Working Paper No. 806.\u003c/li\u003e\n\u003cli\u003eSu, S., Ahmad, A. H., Wood, J., \u0026amp; Jia, S. (2025). Monetary policy analysis using natural language processing: Evaluating the People\u0026rsquo;s Bank of China\u0026rsquo;s minutes and report summary with the Taylor rule. Economic Modelling, 149, 107121. https://doi.org/10.1016/j.econmod.2025.107121\u003c/li\u003e\n\u003cli\u003eWang, Y., \u0026amp; Wang, X. (2024). The role of central bank communication in the long-term stock-bond correlations: Evidence from China. Finance Research Letters, 67, 105893. https://doi.org/10.1016/j.frl.2024.105893\u003c/li\u003e\n\u003cli\u003eYao, J. (2025). A fusion method integrated econometrics and deep learning to improve the interpretability of prediction: Evidence from Chinese carbon emissions forecast based on OLS-CNN model. Computational Economics, 66, 2987\u0026ndash;3006. https://doi.org/10.1007/s10614-024-10793-0\u003c/li\u003e\n\u003cli\u003eYin, M., \u0026amp; Guo, M. (2026). The emotional drive of economic forecasting: GDP predictions based on large language models. Computational Economics. https://doi.org/10.1007/s10614-026-11323-w\u003c/li\u003e\n\u003cli\u003eZarifhonarvar, A. (2026). Generating inflation expectations with large language models. Journal of Monetary Economics, 157, 103859. https://doi.org/10.1016/j.jmoneco.2025.103859\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[{"identity":"75886a62-f44c-479d-aa5c-372ac7781d07","identifier":"10.13039/501100001809","name":"National Natural Science Foundation of China","awardNumber":"62272239","order_by":0},{"identity":"8b77b1cf-7c96-4063-aaf5-ecb0d085a0a9","identifier":"10.13039/501100001809","name":"National Natural Science Foundation of China","awardNumber":"61972208","order_by":1}],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"National Natural Science Foundation of China","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"inflation expectations, China, nowcasting, large language models, explainable artificial intelligence, monetary policy communication","lastPublishedDoi":"10.21203/rs.3.rs-9395947/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9395947/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis paper develops an explainable framework for measuring real-time inflation expectations in China and embedding those signals in inflation nowcasting. The production design is LLM-ready: official texts, news narratives, and search data can be mapped into an economically structured taxonomy of inflation mechanisms. For transparent evaluation, however, the article implements a reduced-form public-data pilot rather than claiming a fully historical LLM deployment. The pilot combines monthly official CPI, PPI, M2 growth, and one-year LPR data for 2023M1–2025M12 with two reproducible expectation channels: a policy-communication index derived from PBOC materials and a lagged attention proxy that approximates public information demand under open-data constraints. In expanding-window nowcasts, expectation-augmented models improve on stricter macro-only linear benchmarks; the best full random-forest specification records an RMSE of 0.253, compared with 0.273 for a macro elastic-net and 0.359 for a simple macro linear model. Feature attribution shows that lagged CPI, PPI, and money growth remain the main anchors, while communication and the lagged attention proxy provide non-trivial incremental information. An exploratory policy-window exercise suggests that the same narrative environment tends to coincide with easing episodes, although that evidence is descriptive because the sample contains only five LPR cuts. The paper’s main contribution is therefore methodological and institutional: it shows how an LLM-ready narrative measurement system can be disciplined by transparent public data, explainable modeling, and cautious validation in China’s recent low-inflation regime.\u003c/p\u003e","manuscriptTitle":"Real-Time Inflation Expectations in China: An Explainable Nowcasting Framework and a Reproducible Communication–Attention Pilot","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-14 09:52:27","doi":"10.21203/rs.3.rs-9395947/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"84a46434-ebd5-48fe-9484-025485346ab8","owner":[],"postedDate":"April 14th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":66163605,"name":"Other Business"}],"tags":[],"updatedAt":"2026-04-14T09:52:27+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-14 09:52:27","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9395947","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9395947","identity":"rs-9395947","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}