‍

For decades, economists and financial professionals have relied on a broad range of indicators to assess the health of the economy and anticipate market movements. While these indicators remain valuable, they all share one major limitation: they often provide only an indirect or delayed reflection of economic reality and are interpreted in highly diverse ways by market participants. As such, the information they convey cannot substitute for the primary driver of short-term price formation — market psychology.

‍

At Lancrey-Javal Technologies, we recognized that the advent of large language models has opened the possibility to quantify this very dimension. By analyzing, at scale, the linguistic patterns underlying professional financial discourse, we can directly measure the collective sentiment, expectations, and fears that shape market behavior in real time. Enabled by our proprietary model, Narval AI, this approach transforms market psychology—from an abstract behavioral notion—into a rigorously measurable financial variable.

‍

‍

Economic Sentiment: A Key Driver of Market Movements

A data point that comes in lower than expected is not always interpreted negatively—particularly in the stock market, where psychology exerts a dominant influence on short-term price formation. This phenomenon, long documented in behavioral finance, has been notably examined by Barberis et al. (1998), who emphasize the role of confirmation bias in financial decision-making. According to this bias, individuals tend to interpret new information not in a purely rational manner, but in ways that confirm or challenge their pre-existing beliefs and expectations.

‍

A telling illustration of this mechanism lies in the interpretation of the unemployment rate—a macroeconomic indicator whose meaning is profoundly ambivalent depending on the surrounding economic context. Drawing on the famous Phillips Curve (1958), which establishes an inverse relationship between inflation and unemployment, investors often perceive an increase in unemployment as a positive signal in an inflationary environment, suggesting that an overheated economy is cooling down and that monetary tightening might soon ease. Conversely, the very same increase in unemployment tends to be interpreted as negative in a recessionary context, as it reinforces fears of economic contraction.

‍

This ambivalence was empirically demonstrated by John H. Boyd, Jian Hu, and Ravi Jagannathan (2005) in their paper “The Stock Market’s Reaction to Unemployment News: Why Bad News Is Usually Good for Stocks.” Their findings show that employment announcements affect stock prices in opposite ways depending on whether the economy is in an expansion or a downturn.

‍

These results underscore a central point: although market psychology has long been recognized as a key driver of short-term price dynamics, no technical means previously allowed for its rigorous and continuous quantification. Until the advent of modern language models capable of analyzing collective discourse at scale, investor sentiment remained an abstract behavioral intuition rather than a measurable financial variable.

‍

‍Narval: A Specialized LLM for the Analysis of Financial Discourse and Market Sentiment

Building on the observations outlined above, and following two years of research at the intersection of artificial intelligence and behavioral finance, we developed in 2025 Narval — a proprietary large language model comprising four billion parameters, specifically trained to understand and analyze the subtleties of economic and financial discussions on X (formerly Twitter).

‍

The choice of X as a primary data source stems from its central role in the circulation of financial information. It has become the social network of choice for finance professionals—78% connect daily according to the French Asset Management Association (AFG)—and recent academic studies (Zeitun et al., 2023; Adams et al., 2023; Angelico et al., 2022) have demonstrated a strong correlation between sentiment expressed on X and subsequent movements in financial markets, credit spreads, monetary policy expectations, and inflation forecasts.

‍

Unlike general-purpose models, Narval was conceived from the outset as a domain-specific architecture oriented toward over-specialization. Its design enables it to analyze with remarkable precision the economic and financial reasoning embedded in professional discourse, offering a nuanced reading of mechanisms and concepts unique to this field. Moreover, Narval captures the tonal dimension of communication—pessimism, negativity, criticism, caution, neutrality, optimism—with a level of consistency and accuracy that exceeds that of generative models.

‍

While generalist systems such as ChatGPT, Claude, or Gemini have achieved remarkable linguistic fluency, they remain unsuitable for specialized financial analysis. Their tendency toward factual imprecision and hallucination renders them unreliable for high-stakes quantitative interpretation. Recent academic findings support this design philosophy: a 2025 joint study by the Universities of Taiwan and Singapore (Financial Named Entity Recognition: How Far Can LLMs Go?, Yi-Te Lu & Yintong Huo) demonstrates that domain-specific BERT-type models significantly outperform generalist LLMs on structured financial tasks.

‍

Today, the advent of large language models marks a turning point in the way financial data is understood and exploited. Narval—as the first artificial intelligence capable of rigorously quantifying market psychology—embodies this paradigm shift, bridging the gap between behavioral intuition and measurable economic intelligence.

‍