Trading with Intelligence: A New Framework for Smarter Financial Markets

Author: Denis Avetisyan

Researchers have developed a novel system that uses advanced artificial intelligence to navigate the complexities of financial trading and optimize investment strategies.

FinRS embodies a risk-sensitive architecture, integrating perceptual inputs with dynamic memory allocation and dual decision agents to achieve reflective updates guided by multi-timescale reward signals, thereby enabling nuanced responses to uncertain environments.

This paper introduces FinRS, a risk-sensitive trading framework leveraging large language models for sequential decision-making, drawdown control, and multi-scale reward optimization in real financial markets.

While large language models demonstrate promising reasoning for financial trading, existing agents often lack integrated risk management, hindering performance in volatile markets. This paper introduces ‘FINRS: A Risk-Sensitive Trading Framework for Real Financial Markets’-a novel approach combining hierarchical analysis, dual-decision agents, and multi-timescale reward to align trading with both profit goals and downside risk constraints. Experiments across multiple stocks demonstrate FinRS achieves superior profitability and stability compared to current state-of-the-art methods. Could this framework represent a crucial step towards more robust and intelligent automated trading systems?

The Inherent Limitations of Conventional Financial Models

Conventional financial modeling relies heavily on assumptions of market efficiency and predictable behavior, often proving inadequate when confronted with the inherent noise and dynamism of real-world trading. These models frequently struggle to incorporate the subtle interplay of investor psychology, geopolitical events, and unforeseen economic shifts, leading to systematic underestimation of risk and limited capacity for adaptation. The reliance on historical data, while valuable, can fail to capture the evolving nature of market correlations and the emergence of novel risk factors. Consequently, traditional approaches often necessitate constant recalibration and expert intervention, hindering their effectiveness in rapidly changing environments and creating opportunities for more agile, data-driven strategies.

Despite the potential of Reinforcement Learning (RL) to navigate financial markets, current agents often face significant hurdles in practical application. These algorithms typically demand vast quantities of historical data to learn effective trading strategies, a limitation when dealing with novel assets or rapidly changing market dynamics. More critically, even after extensive training, RL agents frequently exhibit poor generalization capabilities, meaning their performance deteriorates considerably when confronted with conditions differing from those encountered during the training phase. This susceptibility to unseen scenarios-such as unexpected economic shocks or the introduction of new financial instruments-raises concerns about their reliability and robustness in real-world trading environments, necessitating research into methods that improve data efficiency and enhance adaptability beyond established patterns.

FinRS: A Framework Prioritizing Risk-Adjusted Returns

FinRS is a Large Language Model (LLM)-based framework engineered for automated trading strategies operating across multiple, consecutive decision points. Unlike traditional approaches focused solely on maximizing returns, FinRS explicitly integrates risk mitigation as a core objective throughout the trading process. This is achieved by structuring the framework to evaluate potential trades not only on projected profitability but also on associated risk exposures. The multi-stage sequential design allows FinRS to adapt its investment strategy based on evolving market conditions and the outcomes of previous decisions, facilitating dynamic portfolio adjustments. The framework is intended to address the complexities of real-world trading where both gains and losses must be considered for sustainable performance.

FinRS utilizes a Dual-Decision System to improve the robustness of its trading strategy. This system functionally decouples the determination of investment direction – whether to go long or short on an asset – from the specification of position sizing and risk exposure. By treating these as separate, sequential decisions, FinRS avoids the instability that can arise when directional predictions directly influence trade magnitude. The directional component forecasts market movement, while the quantity/risk adjustment component dynamically scales positions based on current portfolio holdings, estimated volatility, and predefined risk parameters. This separation allows for independent optimization of each decision process, leading to more controlled and stable trading behavior, even in volatile market conditions.

FinRS incorporates Risk-Sensitive Reasoning by continuously evaluating current portfolio positions and real-time market volatility data. This allows the framework to move beyond simple profit maximization and actively modulate trade sizes based on calculated risk exposure. Specifically, position information – including asset allocation and unrealized gains/losses – is used to assess overall portfolio sensitivity. Simultaneously, volatility adjustments, derived from metrics such as historical price fluctuations and implied volatility, are applied to dynamically scale trade quantities. This results in a system that reduces position sizes during periods of high volatility or negative portfolio exposure and increases them during stable or positive conditions, thereby aiming to optimize the risk-adjusted return profile.

Empirical Validation of FinRS Performance

FinRS enhances decision-making through the integration of several analytical techniques. Financial Insight Prompting is utilized to generate actionable insights from financial data. This is complemented by causal chain analysis, which identifies relationships between market events and asset performance. Momentum analysis assesses the rate of asset price changes to identify trends, while probabilistic reasoning assigns likelihoods to different outcomes, allowing for risk assessment and portfolio optimization. The combination of these techniques aims to provide a more nuanced and comprehensive understanding of market dynamics, leading to improved investment strategies.

FinRS employs Rolling-Window Analysis to adapt to changing market conditions by iteratively evaluating data over a defined time horizon, discarding outdated information and focusing on recent trends. This technique facilitates responsiveness to non-stationary market dynamics. Complementing this, the framework integrates real-time market news data, processed through natural language processing, to identify and quantify the impact of external events on asset prices. The system assesses news sentiment and relevance, incorporating this information into its decision-making process alongside historical price data. This combined approach allows FinRS to continuously assess the evolving market landscape and adjust its strategies based on both quantitative and qualitative inputs.

Comparative evaluations demonstrate FinRS’ superior performance against six baseline agents – FinGPT, FinMEM, FinCon, FinAgent, A2C, PPO, and DQN – across a five-stock portfolio. FinRS consistently achieved cumulative returns exceeding 50%. Specifically, the framework generated a 54.99% cumulative return on TSLA stock. This performance was further validated by a Sharpe Ratio of 0.67, indicating a favorable risk-adjusted return profile for the TSLA investment.

Refining Decisions Through Advanced Analytical Methods

The FinRS framework utilizes a Multi-Timescale Reward mechanism to facilitate continuous improvement of trading strategies. This system evaluates performance across varying time horizons, from short-term trade execution to long-term portfolio growth. By assigning rewards based on multiple timescales, the framework can identify strategies that exhibit consistent profitability and adapt to changing market conditions. This approach differs from single-timescale evaluation, which may prioritize immediate gains at the expense of long-term sustainability. The multi-timescale approach enables FinRS to refine its algorithms by weighting rewards based on the relative importance of short-term and long-term objectives, resulting in more robust and adaptable trading behavior.

FinRS utilizes Hierarchical Information Processing (HIP) to manage the complexity of financial data streams. This approach involves processing data in successive layers, beginning with raw market data and progressing through feature extraction, pattern recognition, and ultimately, signal generation. Each layer filters and aggregates information, reducing dimensionality and isolating relevant signals while attenuating noise. Specifically, lower layers focus on identifying short-term fluctuations and technical indicators, while higher layers analyze long-term trends and macroeconomic factors. This layered architecture allows the framework to discern meaningful patterns from high-frequency data and integrate diverse data sources, improving the accuracy and robustness of trading decisions.

FinRS incorporates Conditional Value at Risk (CVaR) and the Kelly Criterion to manage portfolio risk and optimize position sizing. CVaR, a statistical measure of tail risk, quantifies the expected loss given that a certain threshold is exceeded, providing a more comprehensive view of downside exposure than standard Value at Risk. The Kelly Criterion then determines the optimal fraction of capital to allocate to an asset, aiming to maximize long-term growth while minimizing the risk of ruin. Backtesting on TSLA stock revealed a maximum drawdown of 42.34% when utilizing these risk management techniques within the FinRS framework, indicating the system’s performance under adverse market conditions.

The Ascendancy of Intelligent Financial Systems

Financial trading systems are evolving beyond traditional algorithms, and FinRS marks a considerable advancement toward truly intelligent and adaptive platforms. This framework isn’t simply about executing trades faster; it’s engineered to understand the nuanced interplay of market forces. By leveraging sophisticated reasoning capabilities, FinRS actively navigates complex dynamics – accounting for factors beyond simple price movements. Crucially, the system prioritizes enhanced risk management, moving beyond static parameters to dynamically assess and mitigate potential losses. This proactive approach allows FinRS to respond to unforeseen events and changing market conditions with a level of sophistication previously unattainable, promising a future where financial trading is both more profitable and more resilient.

The FinRS framework isn’t a static system; its architecture is deliberately modular, allowing for ongoing refinement and the incorporation of new data sources and algorithms. Crucially, this design integrates Large Language Models (LLMs), transforming the system’s capacity for learning and adaptation. Through LLMs, FinRS can analyze vast quantities of financial news, social media sentiment, and market reports – data previously difficult to quantify – and translate these insights into informed trading decisions. This continuous learning process enables the framework to not simply react to changing market conditions, but to anticipate them, dynamically adjusting strategies and risk parameters. The result is a financial trading system capable of evolving alongside the market, rather than being constrained by pre-programmed rules, ultimately enhancing its resilience and potential for sustained performance.

A critical assessment of the FinRS framework revealed the substantial impact of its risk-sensitive reasoning component on trading performance. When this feature was removed from simulations focused on Tesla (TSLA) stock, the system experienced a marked decline in profitability, with cumulative returns dropping by 15.8 percentage points. More concerningly, the absence of risk-sensitive reasoning led to a dramatic increase in maximum drawdown – a measure of peak-to-trough decline – surging to 82.89%. This finding underscores the component’s vital role in not only identifying profitable opportunities but also in actively mitigating potential losses, demonstrating that intelligent financial trading necessitates a system capable of both maximizing gains and safeguarding capital against adverse market movements.

The framework detailed within this study emphasizes a rigorous approach to sequential decision-making, demanding demonstrable correctness in its algorithms – a principle mirrored in Linus Torvalds’ assertion: “Most good programmers do programming as a hobby, and many of those will eventually realize that they’d rather spend their time writing code than fixing bugs.” The FinRS model, by prioritizing risk-sensitive reasoning and drawdown control, attempts to minimize the need for post-hoc ‘bug fixing’ in trading strategies. It strives for a provably sound methodology, rather than relying on empirical results alone. This parallels Torvalds’ preference for elegant, correct solutions over merely functional ones, focusing on inherent reliability within the system’s design.

What Lies Ahead?

The introduction of FinRS, while a logical progression in applying large language models to financial markets, merely shifts the locus of uncertainty. The framework demonstrably attempts to address drawdown – a statistically inevitable consequence of any trading strategy – but does not fundamentally resolve the inherent unpredictability of the underlying system. True elegance would lie in a provable guarantee of capital preservation, not merely a reduction in the magnitude of loss. The multi-scale reward function, while intuitively appealing, introduces additional parameters demanding rigorous sensitivity analysis – a task often glossed over in pursuit of immediate empirical gains.

A critical, and largely unaddressed, limitation resides in the deterministic reproducibility of results. If a trading strategy, even one guided by a theoretically sound framework, cannot be reliably replicated given identical initial conditions, its value diminishes rapidly. The ‘black box’ nature of large language models exacerbates this issue; understanding why a particular decision was made remains paramount, not simply observing that a decision was made. The field requires a move toward explainable AI, coupled with formal verification techniques.

Future work should prioritize the development of adversarial training methods, specifically designed to test the robustness of FinRS against unforeseen market conditions. Furthermore, exploring the integration of symbolic reasoning – a means of representing and manipulating knowledge in a logically consistent manner – could offer a pathway toward strategies that are not merely reactive to historical data, but capable of genuine, deductive foresight. The pursuit of profit, after all, should not preclude the pursuit of mathematical truth.

Original article: https://arxiv.org/pdf/2511.12599.pdf

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