Can Language Models Predict the Market?

Author: Denis Avetisyan

A new approach combines the reasoning power of artificial intelligence with time-series analysis to generate more accurate and explainable stock forecasts.

The Verbal Technical Analysis framework enhances time-series forecasting by first training a large language model to reason about the data, then leveraging those reasoning outputs to condition the forecasting model—resulting in forecasts accompanied by interpretable reasoning traces.

Researchers introduce Verbal Technical Analysis (VTA), a framework leveraging Large Language Models for improved financial forecasting and interpretability.

Despite advances in financial forecasting, effectively bridging the gap between textual insights and quantitative time-series data remains a significant challenge. This is addressed in ‘Reasoning on Time-Series for Financial Technical Analysis’, which introduces Verbal Technical Analysis (VTA), a novel framework that integrates Large Language Models with time-series modeling. VTA generates accurate and interpretable stock predictions by converting price data into textual annotations and using this reasoning to condition a time-series forecasting model. Could this approach unlock a new era of explainable and high-performing financial forecasting systems?

Beyond the Numbers: Seeking Financial Understanding

Traditional technical analysis, limited by manual interpretation, struggles to scale and introduces bias. While practitioners employ tools like Moving Average Convergence Divergence, Momentum oscillators, and the Relative Strength Index, these often fall short in adapting to the complexities of modern financial markets. A core limitation lies in quantifying qualitative insights – subjective assessments of news, macro trends, and company developments. The pursuit of predictive accuracy often feels elusive, but a truly elegant solution illuminates why something will happen, not simply what.

The VTA framework generates reasoning traces and forecasts, demonstrating its ability to articulate a line of thought.

Verbal Technical Analysis: Giving Voice to the Markets

A novel methodology, Verbal Technical Analysis (VTA), integrates time-series modeling with the reasoning capabilities of Large Language Models (LLMs). This bridges the gap between quantitative data and qualitative interpretation in financial analysis. VTA transforms numerical data into descriptive linguistic statements, enabling LLMs to ‘understand’ underlying behavior. For example, a rising price trend might be annotated as “sustained upward momentum.” By grounding LLM reasoning in quantifiable data, VTA generates interpretable forecasts, addressing the ‘black box’ problem and facilitating a clear audit trail.

Reasoning with Language: The Architecture of Prediction

The core of VTA lies in its Verbal Reasoning component, an LLM trained to identify patterns and predict trends from textual market descriptions. This LLM discerns value signals embedded within descriptive data, enabling quantitative analysis of qualitative information. To optimize predictive capabilities, Conditional Training and Multi-Head Attention mechanisms focus the model on relevant textual features. This nuanced understanding goes beyond simple keyword identification.

A joint conditional training component facilitates the collaborative learning of related concepts.

Further enhancements are achieved through Group Relative Policy Optimization (GRPO) and a time-series adaptation, Time-GRPO. Leveraging an Inverse Mean Squared Error reward function, initial reinforcement learning fine-tuning yielded a 1.6% improvement, followed by a substantial 20.3% gain through rejection sampling and Supervised Fine-Tuning (SFT).

From Forecast to Insight: Empowering Investment Decisions

VTA represents a significant advancement in financial forecasting, delivering interpretable predictions with clear rationales for anticipated fluctuations. This transparency distinguishes VTA from ‘black box’ approaches, allowing analysts to understand why a forecast is generated. The VTA framework seamlessly integrates with existing portfolio management tools, complementing techniques like Markowitz optimization by providing improved input forecasts. Empirical evaluations demonstrate VTA’s superior performance, achieving a state-of-the-art Sharpe Ratio and suggesting an ability to generate substantial risk-adjusted returns.

Expanding the Horizon: The Future of VTA

Future development will focus on integrating diverse data streams, including news articles and social media sentiment, to capture market-moving events and investor psychology. Natural language processing will extract relevant information and quantify its impact. A key expansion involves extending VTA to multi-asset portfolios, implementing algorithms for efficient construction and risk allocation, and adapting to changing market conditions. The ultimate goal is to establish VTA as a comprehensive, AI-powered platform, consistently achieving high ratings in Clarity, Depth, Accuracy, Coherence, and Relevance.

The pursuit of insightful financial forecasting, as demonstrated in this framework, necessitates a shift in perspective—a reimagining of how data is interpreted. This echoes Thomas Kuhn’s observation: “The world does not speak to us directly, it is we who speak to it.” The Verbal Technical Analysis (VTA) framework embodies this principle by translating raw time-series data into natural language, allowing for a more nuanced and human-understandable analysis. This isn’t merely about achieving higher prediction accuracy; it’s about constructing a system that reasons about financial markets, mirroring the cognitive processes of experienced analysts. The elegance of VTA lies in its ability to bridge the gap between quantitative data and qualitative understanding, creating a holistic approach to forecasting that prioritizes both precision and interpretability.

What’s Next?

The pursuit of predictive accuracy in financial markets often feels like chasing a phantom. This work, by attempting to bridge the gap between the rigor of quantitative time-series analysis and the nuanced, if often subjective, world of verbal reasoning, offers a subtle shift in perspective. However, the very success of Verbal Technical Analysis (VTA) highlights a persistent question: how much of financial ‘insight’ is genuinely signal, and how much is merely a compelling narrative constructed post hoc? The framework’s reliance on Large Language Models, while powerful, underscores the inherent challenge of extracting objective truth from inherently probabilistic systems.

Future iterations should address the limitations of current LLM’s in handling genuinely novel market conditions – those events that fall outside the scope of their training data. Reinforcement learning, as a component of VTA, presents a particularly fertile ground for exploration; moving beyond simple reward functions to incorporate measures of ‘cognitive consistency’ within the LLM’s reasoning process could prove illuminating. The true test will not be simply achieving higher prediction scores, but demonstrating a capacity for graceful degradation when faced with unforeseen circumstances.

Ultimately, elegance isn’t optional; it is a sign of deep understanding and harmony between form and function. A system that can articulate not just what will happen, but why, and can do so with a minimum of complexity, is not only more useful, but more durable and comprehensible. Beauty and consistency make a system durable and comprehensible.

Original article: https://arxiv.org/pdf/2511.08616.pdf

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