Words That Move Markets: How Earnings Calls Shape Expectations

Author: Denis Avetisyan

New research reveals that the narrative framing of corporate earnings calls-beyond the numbers themselves-systematically influences analyst forecasts and ultimately, realized earnings.

Analyst forecasts demonstrate a pronounced clustering immediately before earnings calls, suggesting a strategic, yet potentially biased, pattern of information dissemination rather than continuous, independent assessment of company performance-a behavior quantified by the concentration of predictions within a narrow temporal window $t$.

This study utilizes textual analysis and counterfactual modeling to demonstrate the impact of linguistic cues in earnings calls on analyst beliefs and financial outcomes.

While quantitative data dominates financial analysis, the narratives framing that data powerfully shape investor expectations. This is the central question explored in ‘Corporate Earnings Calls and Analyst Beliefs’, which demonstrates that linguistic framing-specifically the narratives conveyed during earnings calls-significantly influences analyst forecasts and realized earnings. By employing a novel text-morphing methodology with large language models, the study reveals systematic biases in how analysts react to optimism versus risk. Ultimately, this research asks whether a more granular understanding of narrative’s impact can refine our models of expectation formation and improve market efficiency.

The Illusion of Objectivity: Beyond Numbers in Financial Assessment

For decades, financial analysis has been fundamentally rooted in the meticulous examination of quantitative data – balance sheets, income statements, and key performance indicators. However, this emphasis often overshadows the potent, yet less tangible, influence of qualitative narratives. Financial markets aren’t driven solely by numbers; they are profoundly shaped by the stories companies tell about themselves, their strategies, and their future prospects. These narratives, communicated through various channels, create expectations and influence investor sentiment in ways that pure data rarely can. While quantitative analysis provides a crucial foundation, a comprehensive understanding of financial health requires acknowledging that the perception of financial reality, as constructed through compelling storytelling, is often just as critical as the reality itself. Ignoring this narrative component can lead to inaccurate valuations and miscalculated risks, underscoring the need for a more holistic approach to financial assessment.

Financial analysts routinely integrate both hard numerical data and the softer influence of corporate narratives when formulating expectations about a company’s future performance. While quantitative analysis provides a baseline understanding of financials, these narratives – delivered through avenues like earnings calls and investor presentations – frequently act as crucial signals, capable of reinforcing or even overriding purely data-driven conclusions. However, a significant hurdle exists in systematically evaluating the impact of these narratives; their nuanced and often subjective nature resists easy quantification. Researchers are actively developing methods-from natural language processing to sentiment analysis-to translate narrative content into measurable variables, aiming to determine the extent to which specific linguistic patterns, emotional tones, or framing choices actually shift analyst forecasts and, ultimately, market valuations. The difficulty lies not just in identifying relevant narrative elements, but also in isolating their independent effect from the constant flow of quantitative information and broader market trends.

Research demonstrates that the narratives presented during corporate Earnings Calls wield a surprising degree of influence over financial analysts, often surpassing the impact of strictly numerical data. These calls aren’t merely reports of performance; they are carefully constructed stories designed to frame results and shape future expectations. A compelling narrative can recontextualize even disappointing figures, fostering optimism and maintaining positive analyst projections. Conversely, even strong quantitative results can be undermined by a poorly articulated or negative narrative. This phenomenon suggests that analysts aren’t purely rational actors, but are susceptible to framing effects and the persuasive power of storytelling, highlighting the critical need to analyze how information is conveyed, not just what information is presented.

Accurate financial forecasting and robust risk assessment increasingly demand a synthesis of quantitative data with the nuanced understanding of prevailing narratives. While historical financial performance provides a baseline, investor expectations – powerfully shaped by stories communicated through sources like earnings calls – often dictate market reactions. Ignoring these narratives introduces a systematic bias, potentially leading to mispriced assets and underestimated risks. Research demonstrates that the impact of these qualitative signals can, at times, outweigh the influence of purely numerical data, highlighting the necessity for models that incorporate textual analysis and sentiment evaluation. By acknowledging and quantifying this interplay between data and narrative, financial professionals can move beyond simplistic projections and build more resilient, informed strategies.

Analysts' beliefs and ultimately realized earnings are significantly influenced by the narratives presented to them. — Analysts’ beliefs and ultimately realized earnings are significantly influenced by the narratives presented to them.

From Text to Signal: Encoding Narrative with Modern Techniques

Textual embeddings transform narrative text into numerical vectors, capturing semantic meaning for computational processing. This process involves mapping words, phrases, or entire documents into a high-dimensional vector space where proximity between vectors indicates semantic similarity. Algorithms like Word2Vec, GloVe, and more recently, those utilized within Large Language Models, are employed to create these embeddings. The resulting vectors allow for quantitative analysis of textual data, enabling tasks such as sentiment analysis, topic modeling, and document similarity comparisons. Unlike traditional bag-of-words models, textual embeddings preserve contextual information, representing the meaning of a word based on its surrounding text, and allowing for a more nuanced understanding of the narrative content.

Large Language Models (LLMs), including Llama 3 and FinBERT, are utilized to create textual embeddings from Earnings Calls by processing the transcripts and converting the language into numerical vector representations. These embeddings capture semantic information, allowing for the quantification of narrative sentiment expressed during the call. FinBERT, specifically trained on financial text, excels at identifying sentiment relevant to financial performance, while general-purpose LLMs like Llama 3 can be adapted for this purpose. The resulting numerical data enables analysts to move beyond simple keyword counts and assess the overall emotional tone and subtle cues within the earnings narrative, providing a more granular understanding of management’s outlook and potential market impact.

Traditional natural language processing (NLP) techniques, such as bag-of-words or term frequency-inverse document frequency (TF-IDF), often fail to capture contextual relationships and semantic subtleties within text. Large Language Models (LLMs), conversely, utilize transformer architectures and are pre-trained on massive datasets, enabling them to understand word order, polysemy, and implicit meaning. This allows LLMs to discern nuances like sarcasm, hedging, and sentiment shifts that are missed by methods relying solely on keyword frequency or predefined lexicons. Consequently, LLM-generated textual embeddings provide a richer and more accurate representation of narrative influence, facilitating improved analysis of factors like investor sentiment and the impact of specific statements on financial forecasts.

Counterfactual analysis, enabled by Large Language Models (LLMs), involves systematically altering the textual content of narratives – such as earnings call transcripts – and then observing the resulting impact on quantitative forecasts. This process allows researchers and analysts to move beyond simply identifying correlations between narrative and outcomes, and instead assess causal relationships. For example, specific phrases expressing confidence or caution can be substituted with alternatives, and the LLM re-processes the modified text to generate a new forecast. The difference between the original forecast and the revised forecast provides a quantifiable measure of the phrase’s influence. This technique requires careful control for confounding variables and robust statistical analysis to determine the significance of observed changes, but offers a powerful method for understanding narrative effects on financial predictions and other modeled outcomes.

Narratives significantly influence the degree of disagreement among forecasts.

Beyond Correlation: Validating Narrative Impact with Statistical Rigor

Integration of Large Language Model (LLM)-generated narrative embeddings into forecasting models demonstrates a significant contribution of textual data to predictive accuracy. Specifically, narrative features account for 25-35% of the total predictive power observed in these models. This performance is statistically comparable to that achieved using traditional fundamental factors, such as financial ratios and economic indicators. The methodology involves converting textual data-including news articles, earnings call transcripts, and social media posts-into numerical vector representations (embeddings) which are then incorporated as independent variables in time-series forecasting models. This indicates that market-relevant information conveyed through narrative sources is a substantial driver of predictive outcomes and complements existing quantitative analysis.

The Clark-West test is a statistical procedure employed to evaluate the economic significance of incorporating narrative data into predictive models. This test assesses whether the incremental explanatory power gained from including narrative features – as measured by the difference in model performance metrics like R-squared or Root Mean Squared Error – is statistically different from zero. Specifically, the test examines the null hypothesis that the coefficients on the narrative variables are jointly equal to zero, using an F-statistic and corresponding p-value to determine statistical significance. A statistically significant p-value, typically below 0.05, indicates that the narrative data contributes meaningfully to the model’s predictive accuracy beyond what is already explained by fundamental factors, thereby validating its informational value.

Analysis of Surprise Unexpected Earnings (SUE) demonstrates the predictive capability of narrative data regarding market responses to earnings announcements. Specifically, models incorporating LLM-generated narrative embeddings show a statistically significant correlation between textual sentiment surrounding earnings and subsequent stock price movements following the SUE event. This indicates that narratives capture information not fully reflected in traditional financial data, allowing for anticipation of market reactions. The degree of correlation varies by sector, but consistently demonstrates that narrative analysis contributes meaningfully to understanding and potentially forecasting post-earnings announcement price volatility and direction.

Analysis of analyst forecast disagreement, when correlated with narrative data, indicates that narratives focused on uncertainty exhibit the most significant impact on forecast dispersion. Specifically, higher levels of textual uncertainty, as measured through LLM-generated embeddings, directly correlate with increased variance in analyst earnings estimates. This suggests that when narratives emphasize ambiguity or potential volatility, analysts demonstrate a wider range of predictions, reflecting the inherent difficulty in projecting outcomes under uncertain conditions. The strength of this correlation provides empirical evidence that narrative-driven uncertainty is a key driver of forecast disagreement, comparable to, and often exceeding, the influence of traditional financial metrics.

Forecast disagreement generally increases with prediction horizon, indicating decreasing confidence in longer-term predictions.

The Future is Nuance: Embracing Narrative in Financial Prediction

Financial forecasting traditionally relies on quantitative data, yet often overlooks the powerful influence of qualitative information – specifically, the narratives surrounding companies and markets. Recent research demonstrates that integrating narrative analysis into financial models significantly enhances predictive accuracy and reduces investment risk. By examining the language used in earnings calls, news reports, and regulatory filings, analysts can gain deeper insights into management sentiment, perceived opportunities, and potential threats. This approach moves beyond simply what is reported, to understanding how it is communicated, revealing subtle cues often missed by conventional methods. The resulting models are more robust, capable of anticipating market reactions with greater precision, and ultimately leading to more informed and successful investment decisions by accounting for the complex interplay between information and interpretation.

The efficacy of financial markets hinges not solely on quantitative data, but also on the qualitative influence of narratives that shape analyst expectations. Research demonstrates that the stories surrounding a company – whether emphasizing growth potential or highlighting emerging risks – significantly impact forecasting and, consequently, market reactions. By systematically analyzing the narratives present in earnings calls, news reports, and other communications, it becomes possible to anticipate how analysts will interpret information and adjust their recommendations. This insight enables the development of more informed trading strategies, allowing investors to capitalize on predictable biases or preemptively adjust portfolios in response to evolving sentiment. Ultimately, understanding the narrative landscape provides a crucial edge in navigating the complexities of financial markets and mitigating the risks associated with unforeseen shifts in investor perception.

Research indicates a compelling relationship between the narrative tone surrounding financial reports and subsequent outcomes. Specifically, reports framed with confidence-emphasizing positive outlooks and assured projections-demonstrate a significant positive correlation with both analyst forecasts and ultimately, realized earnings. Conversely, narratives steeped in uncertainty, highlighting potential risks and ambiguous conditions, are linked to a pronounced negative impact on market performance, suggesting that analysts consistently underemphasize the implications of disclosed uncertainties. This pattern reveals a systematic bias in how financial information is processed, where optimistic storytelling can amplify positive expectations and, potentially, results, while cautious disclosures fail to adequately temper projections, leading to market corrections.

The principles of narrative-informed prediction, initially demonstrated through analysis of earnings forecasts, demonstrate a broader applicability across the financial landscape. Beyond simply predicting company performance, this methodology offers potential for enhanced credit risk assessment by gauging the prevailing narratives surrounding borrower stability and future repayment capacity. Furthermore, macroeconomic modeling stands to benefit, as the analysis of national and global narratives – concerning consumer confidence, geopolitical stability, or technological disruption – can provide leading indicators often missed by traditional quantitative data. This suggests a future where financial modeling isn’t solely reliant on numbers, but actively incorporates the power of story to anticipate shifts in market sentiment and overall economic trends, ultimately leading to more resilient and insightful financial strategies.

The study illuminates how seemingly immaterial linguistic framing during corporate earnings calls systematically alters analyst forecasts-a phenomenon readily echoing Newton’s observation that, “If I have seen further it is by standing on the shoulders of giants.” The ‘giants’ here aren’t predecessors in physics, but the accumulated biases and interpretive frameworks analysts bring to bear. The paper demonstrates that narratives, much like initial conditions in a complex system, can exert a disproportionate influence on outcomes, even when quantitative data remains constant. It’s a potent reminder that a model isn’t a mirror of reality-it’s a mirror of its maker, reflecting pre-existing beliefs as much as objective truth. The significance level of this narrative effect warrants careful consideration.

Where Do We Go From Here?

The demonstrated susceptibility of analyst forecasts to linguistic framing begs a question rarely asked with sufficient rigor: how much of financial ‘analysis’ is simply sophisticated storytelling? This work does not disprove quantitative models, but it does illustrate their inherent fragility when confronted with the messiness of narrative. The field now faces the challenge of quantifying that error – not merely refining the models themselves. One anticipates a surge in attempts to build ‘narrative risk’ metrics, though the temptation to reduce complex communication to a set of easily measurable keywords should be viewed with skepticism.

A critical limitation lies in disentangling correlation from causation. While earnings call narratives demonstrably precede forecast revisions, proving a direct causal link remains elusive. Future research should embrace more robust counterfactual designs, perhaps leveraging synthetic control groups generated through large language models. However, even perfect causal inference will not address the fundamental issue: the narratives themselves are not exogenous shocks, but emergent properties of complex organizational behavior.

Ultimately, the pursuit of predictive accuracy may be a distraction. A more fruitful path lies in understanding why these narratives are effective – what cognitive biases do they exploit, and how do they shape the subjective probabilities that drive investment decisions? Wisdom, it seems, is not knowing the future, but knowing your margin of error-and appreciating just how easily that margin can be manipulated.

Original article: https://arxiv.org/pdf/2511.15214.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

The Illusion of Objectivity: Beyond Numbers in Financial Assessment

From Text to Signal: Encoding Narrative with Modern Techniques

Beyond Correlation: Validating Narrative Impact with Statistical Rigor

The Future is Nuance: Embracing Narrative in Financial Prediction

Where Do We Go From Here?

See also: