Causal Economic Narratives

• Causal economic narratives are structured chains of cause–effect propositions extracted from discourse, using explicit linguistic cues and semantic similarity.
• They integrate methods like rule-based detection, transformer embeddings, and statistical indices to quantify and map narrative impacts across markets and policies.
• Applications include forecasting economic turning points, simulating market dynamics in agent-based models, and evaluating policy narratives through empirical indices.

Causal economic narratives are structured chains of cause–effect propositions, extracted from economic discourse—text, survey responses, or policy statements—that explicitly or implicitly attribute economic events to identifiable drivers. These narratives function both as objects of empirical measurement (via formal textual extraction pipelines and statistical indices) and as agents of economic behavior and policy formation, with consequential feedback loops between economic outcomes and the stories that describe them. The domain covers methodologies for extraction and quantification, formal equilibrium models of narrative-driven beliefs and choices, agent-based simulations of narrative-driven markets, and empirical studies of the bidirectional causal impact of narratives and economic variables.

1. Formalization and Extraction of Causal Economic Narratives

The structural representation of causal economic narratives begins at the sentence or clause level. A causal sentence is defined as having a “cause” clause $X$ and an “effect” clause $Y$, linked by explicit cue expressions (“because of,” “due to,” “therefore,” etc.). Extraction pipelines identify these pairs using syntactic and morphological parsing alongside rule-based detection of 41+ cue patterns, as implemented in Japanese economic survey analysis and climate-change narrative mining (Shigetsugu et al., 2024, Sakaji et al., 2024).

At higher levels, a causal chain (or narrative) is constructed when an effect clause $Y$ in topic $A$ (at time $t-d$) is semantically similar to a cause clause $Y'$ in topic $B$ (at time $t$), establishing a transitive link $X \to Z$ across topics and time. Both (Shigetsugu et al., 2024) and (Sakaji et al., 2024) utilize transformer-based embeddings (Japanese Sentence-BERT or Financial BERT) to map text spans into a vector space, and link spans when cosine similarity exceeds a threshold (e.g., 0.5 or 0.7).

Beyond survey-based approaches, “Causal Micro-Narratives” (Heddaya et al., 2024) operationalizes sentence-level explanations for a fixed economic subject (e.g., inflation) using a subject-specific ontology of possible cause and effect categories. Each sentence is classified as to whether it contains a narrative, and further tagged with applicable causes/effects through fine-tuned LLMs, yielding high F1 scores on large-scale annotated datasets.

Table: Core Components of Causal Economic Narrative Extraction

Component	Details	Papers
Low-level structure	(Cause, Effect) clauses linked by cues; span embeddings	(Shigetsugu et al., 2024, Sakaji et al., 2024, Heddaya et al., 2024)
Cross-topic/time link	Transitive linking via semantic similarity (Sentence-BERT/BERT), usually cosine > 0.5 or 0.7	(Shigetsugu et al., 2024, Sakaji et al., 2024)
Ontological labels	Domain-specific categories (e.g., causes/effects for inflation); multi-label annotation	(Heddaya et al., 2024)
Aggregation	Narrative indices as time series over topic pairs or categories	(Shigetsugu et al., 2024, Sakaji et al., 2024, Heddaya et al., 2024)

2. Quantification, Aggregation, and Index Construction

Quantifying economic narratives relies on aggregating causal chains across time and topic dimensions. In the “Keiki Watchers Survey” analysis, each topic pair $(A,B)$ forms a narrative index: $$\text{Index}_{A\to B}(m) = \sum_{(i,j,t,d)\in \mathcal{C}_{A\to B}(m)} w(\vec{i}_{t-d},\vec{j}_t,d)$$ where the weight $w(\cdot)$ applies a decaying function of semantic similarity and time: $$w(\vec{i},\vec{j},d) = \frac{1}{1 + a\,e^{b\,d} \cos(\vec{i},\vec{j})}$$ with $a, b$ calibrated so weights halve over a target time frame (e.g., 5 years) (Shigetsugu et al., 2024).

In the climate-change context, indices aggregate strengths of all causal chains from one topic to another within a time window, forming a high-dimensional time series (e.g., a 1,560-dimensional vector for 40 topics) (Sakaji et al., 2024). These indices can be visualized as temporal plots or cause–effect network graphs, and clusterable via techniques such as PCA or $k$-means.

Empirical performance is benchmarked via correlation (e.g., Pearson’s $\rho$) against standard business-cycle indicators (leading/lagging diffusion indices). Key findings include robust tracking of cumulative lagging DIs and tight alignment between specific narrative chains and economic turning points, with heat-maps revealing that certain causal paths systematically anticipate or coincide with macroeconomic shifts (Shigetsugu et al., 2024).

3. Mechanisms and Effects: Equilibrium, Belief, and Agent-Based Models

Beyond text extraction, formal models treat economic narratives as explicit causal maps that mediate economic beliefs, behaviors, and policy equilibria.

Eliaz and Spiegler (Eliaz et al., 2018) formalize narratives as Bayesian networks ($R$) mapping from policy choices ($a$) to consequences ($y$), with intermediate variables. Each agent’s anticipatory utility arises not from objective accuracy but from "hopeful" causal stories, yielding equilibrium as a steady-state over narrative–policy pairs. Feedback loops are intrinsic: prevailing narrative–policy pairs alter the perceived relationship between policies and outcomes, which in turn affect policy choice distributions. Theoretical results show:

• Only two policies obtain positive equilibrium mass under “perfect” narrative DAGs.
• Narrative polarization and drift emerge naturally as equilibrium dynamics.
• Non-linear/opportunity narratives can sustain extreme, polarized policy mixtures.

Agent-based models (Lomas et al., 2020) instantiate narrative economics in artificial financial markets. Here, "narratives" are agent-level expectations (opinions) about future asset value, summarized as scalar beliefs $x_i \in [-1,1]$. Opinion dynamics (bounded-confidence, relative-agreement, relative-disagreement) govern narrative spread, while trader decision-rules (OZIC, ONZI) directly embed these beliefs into quote formation. Simulations demonstrate that:

• Extreme optimism/pessimism in narratives produces large, persistent price bubbles or crashes.
• Exogenous narrative shocks (opinion flips) cause abrupt discontinuities in asset prices.
• The entire price dynamics can be causally traced to the propagation and revision of agent narratives.

Such models confirm that narrative-driven beliefs, even among low-intelligence agents, suffice to generate empirically relevant volatility, clustering, and regime switches (Lomas et al., 2020).

4. Empirical Causality Between Narratives and Economic Outcomes

Recent work empirically quantifies the two-way causal linkages between economic news narratives and market outcomes. Using high-frequency financial data and large-scale news corpora, “Stories that (are) Move(d by) Markets” (Drinkall et al., 20 Feb 2025) measures the semantic shift in news as the day-to-day cosine distance of aggregated transformer embeddings. Causal identification exploits:

• Structural financial “shocks” extracted via SVAR models with sign restrictions (growth, monetary, common premium, hedging premium).
• Granger-causality analyses to establish whether semantic shifts in text predict (or are predicted by) market shocks, controlling for historical lags.

Key findings include:

• Textual semantic shifts are Granger-caused by market shocks, with up to 25% significant Granger-causal pairs for left-leaning outlets (lower for right-leaning).
• In reverse, text-to-market causality appears at longer lags (5–10 days) and is stronger during highly exogenous events such as COVID-19, where text-based signals enhance market shock forecasting by up to 15–20% in relative MSE reduction.
• Statistically significant bidirectional feedback loops, modulated by outlet partisanship, are documented for certain classes of economic shocks.
• Controls for confounders and robustness checks confirm that the findings are resilient to linear/non-linear model choice and multiple-testing corrections.

This establishes that narrative formations (and their semantic shifts) both respond to, and can predict, structural market changes in real time, revealing a dynamically coupled narrative–market system (Drinkall et al., 20 Feb 2025).

5. Policy Narratives and Dynamic Institutionalization

The institutional transmission of narratives is formalized in dynamic game-theoretic frameworks. The “Narratives–Construct–Commitment” (NCC) model (Jiang et al., 29 Apr 2025) examines how government-originated public narratives first shape private expectations (Construct stage), and, through reinforcement and local implementation, become institutional commitments sustaining economic growth.

Key state variables include:

• Aggregate credibility belief ($O_t$)
• Accumulated institutionalization ($L_t$)
• Share of market skeptics ($w_t$)

The central government designs narrative “precision” ($p_t$) and “monitoring” ($m_t$), with local governments selecting policy consistency ($c_t$), and market participants updating beliefs and investment levels accordingly. Bayesian updating and Markov-perfect equilibrium characterize the evolution.

Empirical tests (local projection on high-frequency financial data) quantify narrative shocks and trace real-world commitment phases (e.g., China's Innovation-Driven Development Strategy), connecting visible narrative events to sustained productivity growth through evolving belief and investment channels. Comparative statics reveal that raising narrative precision or monitoring intensity increases local implementation; heterogeneity in local capability reduces the average effect but can be offset with stronger monitoring (Jiang et al., 29 Apr 2025).

6. Applications, Challenges, and Future Directions

Causal economic narratives have been leveraged for:

• Tracking macroeconomic turning points and forecasting diffusion indices with stronger correlation than traditional diffusion indices (Shigetsugu et al., 2024).
• Analyzing climate transition and physical risk narratives, as well as links from international conferences to national policy formation (Sakaji et al., 2024).
• Real-time financial prediction and risk monitoring, especially during exogenous shocks (e.g., COVID-19) (Drinkall et al., 20 Feb 2025).

Major methodological challenges include ontology design and annotation (for micro-narratives (Heddaya et al., 2024)), the detection of implicit or multi-sentence narratives, and the validation of narrative indices in predictive, not just correlative, terms. Extensions under active development include integration with formal causal inference (e.g., statistical Granger tests), coupling of narrative indices with high-frequency market data for nowcasting, and systematic mapping of narrative feedback loops in complex agent-based or game-theoretic models.

A plausible implication is that as causal narrative extraction and modeling becomes more refined, it will enable not only improved forecasting but also more granular, automated diagnostics of how economic stories propagate, mutate, and crystallize into both public belief and policy—closing the empirical–theoretical–computational narrative economics loop.