Theoretical Foundations, Measurement Methods, and Practical Applications
Joerg Osterrieder
Download PDF (44 pages)Key definitions establishing the mathematical framework for narrative finance.
No formal definitions extracted.
Mathematical results establishing properties of narrative dynamics.
No theorems extracted.
All 15 figures from the primer, generated programmatically for reproducibility.
Overview of all sections in the primer.
The role of narratives in financial markets has gained significant academic attention following Robert Shiller's influential work on narrative economics [shiller2017narrative,shiller2019narrative]. Shiller argues that economic fluctuations are substantially driven by contagious stories that spread t...
We make three principal contributions to the narrative finance literature: Contribution 1: Formal Definition. We propose that a financial narrative is a 5-tuple = (S, C, A, T, E) where: itemize[noitemsep] S = Story: A causally connected sequence o...
Narrative finance occupies a distinctive position relative to two established paradigms in financial economics (Figure~fig:paradigms): figure[htbp] p03_nf_bf_emh_relationship/chart.pdf Relationship between Efficient Market Hypothesis (EMH), Behavioral Finance (BF), and Narrative Finance (NF). Each p...
This primer is organized as follows: itemize[noitemsep] Section~sec:related: Related Work (topic modeling, NLP in finance, narrative economics) Section~sec:theory: Theoretical Framework (definition, taxonomy, contagion models) Section~sec:methods: Methods Landscape (LDA, BERTopic, LLMs, econometric ...
...
...
...
The term ``narrative'' is used loosely across the social sciences. In financial contexts, it is often conflated with ``news,'' ``sentiment,'' or ``themes.'' We propose a rigorous definition that distinguishes narratives from these related concepts. definition[Financial Narrative] def:narrative A
subsec:taxonomy Building on our definition, we propose a two-tier taxonomy distinguishing eight narrative types (Figure~fig:taxonomy). This taxonomy serves both theoretical and practical purposes: theoretically, it organizes the heterogeneous landscape of financial narratives into coherent categorie...
subsec:contagion A distinctive feature of narratives is their epidemic-like spread through populations. We model this using the SIR (Susceptible-Infected-Recovered) framework [shiller2017narrative]. definition[SIR Model for Narratives] Let S(t), I(t), and R(t) ...
subsec:cascades A related phenomenon is the information cascade, where agents rationally ignore their private information and follow the observed actions of others [bikhchandani1992theory]. definition[Information Cascade] def:cascade An information cascade occurs when an ag...
subsec:gap A central methodological challenge is the distinction between topics (as extracted by topic models) and narratives (as defined in Definition~def:narrative). proposition[Topic-Narrative Gap] prop:gap Let T denote the set of topics extracted from corpus ...
We formalize narrative quantification as follows: definition[Narrative Quantification Problem] def:problem Given: itemize[noitemsep] Corpus D = \{d_1, , d_n\} with timestamps \{t_1, , t_n\} Financial returns r = (r_1, , r_T) for T periods itemize F...
Table~tab:methods in Section~sec:related summarizes the method landscape. We organize methods into three generations: enumerate[label=(*)] Traditional (1999--2015): LDA, NMF---probabilistic and algebraic approaches operating on bag-of-words representations Neural (2...
...
Neural methods leverage pre-trained language models to capture semantic similarity beyond lexical matching....
Large language models have transformed NLP since 2023, enabling new approaches to narrative extraction....
From topic model outputs, we construct quantitative indicators capturing different aspects of the narrative definition....
Narrative indicators enter standard econometric models for causal analysis and forecasting....
We have surveyed methods spanning three generations of topic modeling, from LDA (2003) to LLMs (2023+). Key trade-offs include: itemize[noitemsep] Coherence vs. Speed: Neural methods (BERTopic, FASTopic) produce more coherent topics but require embedding computation Depth vs...
...
We present results from the BERTopic pipeline applied to our synthetic corpus. The evaluation focuses on two aspects: static quality (do extracted topics correspond to coherent, ground-truth categories?) and dynamic patterns (do topic prevalences evolve realistically over time?)....
We now examine whether the extracted narrative indicators contain information relevant for understanding and forecasting financial returns. Three complementary analyses are presented: correlation analysis to establish basic relationships, Granger causality tests to assess predictive content, and imp...
We compare out-of-sample forecasting performance using an 80/20 train/test split. The training period comprises months 1-48, and the test period comprises months 49-60. This temporal split ensures that we evaluate true out-of-sample performance without look-ahead bias. Three models are compared to i...
Ablation studies systematically vary components of the pipeline to understand their individual contributions and assess robustness to hyperparameter choices. A reliable methodology should produce consistent results across reasonable parameter ranges; excessive sensitivity to specific choices would u...
All experiments in this primer are designed for full reproducibility. The complete codebase is available at https://github.com/Digital-AI-Finance/Primer_Narrative_Finance under an open-source license. All stochastic components use a single GLOBAL\_SEED = 42 to ensure deterministic outputs across run...
Narrative analysis has proven valuable for understanding historical asset bubbles, where the role of shared beliefs is central. Traditional financial economics struggles to explain why sophisticated investors participate in bubbles they recognize as such; narrative finance provides a framework where...
The cryptocurrency space is particularly narrative-driven, given the absence of traditional fundamentals. Unlike equities with earnings or bonds with cash flows, cryptocurrencies have no intrinsic value proposition beyond what market participants collectively believe. This makes narrative analysis n...
The GameStop short squeeze of January 2021 represents a paradigmatic case of narrative-driven market dynamics. Unlike historical bubbles that unfolded over years, the GME episode compressed narrative formation, collective action, and market impact into a three-week period, providing an unprecedented...
Central bank communication provides a natural laboratory for narrative effects, given the explicit focus on managing expectations. Unlike other market narratives that emerge organically from distributed agents, central bank narratives have identifiable authors (the FOMC, ECB Governing Council, etc.)...
Environmental, social, and governance (ESG) investing is fundamentally driven by narratives about corporate responsibility, climate risk, and sustainable development. Unlike traditional financial narratives that focus on price appreciation or income, ESG narratives introduce moral and ethical dimens...
Narrative finance has significant implications for financial regulation and policy. The recognition that narratives influence market dynamics independent of fundamentals creates both opportunities for more effective policy design and challenges for traditional regulatory frameworks built around info...
We acknowledge several limitations of our approach, organized into data, methodological, and interpretive categories. These limitations should inform researchers about the boundaries of current methods and guide future improvements....
We identify several promising directions for extending the methods and applications presented in this primer....
We frame several open problems as research questions: enumerate[label=Q*.] Identification. How can we establish causal effects of narratives on markets in the presence of simultaneity and omitted variables? What natural experiments or instrumental variables are avai...