Evolving Smarter Traders: How AI is Rewriting the Rules of Finance

Author: Denis Avetisyan

Researchers are leveraging the power of artificial intelligence to automatically design and refine trading strategies, pushing the boundaries of algorithmic finance.

The MadEvolve system iteratively refines programs through a closed-loop process-sampling parent and inspiration code from a population database, querying a large language model ensemble, evaluating candidates with a backtester, and subsequently updating the population-thereby enacting a form of automated, evolutionary optimization, as detailed in Liet al.(2026).

This paper introduces MadEvolve, a system employing large language model-driven evolutionary search to optimize trading systems, demonstrating improved performance and robustness compared to traditional methods.

Despite advances in quantitative finance, optimizing complex trading strategies remains a challenging, high-dimensional search problem. This paper introduces ‘MadEvolve: Evolutionary Optimization of Trading Systems with Large Language Models’, a framework leveraging large language models to evolve algorithmic trading strategies and alpha generation, inspired by recent progress in computational cosmology. We demonstrate significant performance improvements across various tasks-from feature engineering to execution optimization-using a simulation and backtesting setup for Bitcoin trading, while carefully addressing concerns regarding statistical significance and generalization. Could this LLM-driven, agentic approach herald a new era of automated strategy discovery in financial markets?

Navigating the Evolving Landscape of Alpha Generation

Conventional alpha forecasting techniques, historically reliant on static models and pre-defined parameters, are increasingly challenged by the intricacies of contemporary financial markets. These methods often fail to adequately account for the escalating volume of data, the heightened speed of transactions, and the unpredictable nature of market responses. The inherent limitations of static approaches manifest as an inability to adapt to shifting market dynamics, leading to diminished predictive power and reduced profitability. Consequently, strategies built upon these foundations struggle to identify genuine alpha signals amidst the pervasive noise, requiring a fundamental reassessment of forecasting methodologies to maintain competitive performance.

Conventional investment strategies, built upon historically static models, increasingly falter when confronted with the volatile and interconnected nature of contemporary financial markets. These limitations stem from an inability to effectively process the sheer volume of data and adapt to rapidly changing market dynamics, leading to diminished predictive power and suboptimal returns. Consequently, there’s a growing imperative for investment approaches that incorporate dynamic learning and optimization techniques. Such strategies leverage algorithms capable of continuously refining their parameters and adapting to new information, effectively ‘learning’ from market behavior and improving performance over time. This proactive adjustment, unlike the rigidity of traditional methods, allows for a more nuanced and responsive investment process, ultimately seeking to capitalize on emerging opportunities and mitigate potential risks in an ever-shifting landscape.

Modern financial markets demand algorithmic frameworks that transcend static modeling, necessitating a co-evolutionary approach to both feature selection and strategic implementation. This dynamic interplay allows systems to adapt to changing market conditions, continuously refining the inputs used for decision-making alongside the strategies themselves. Research demonstrates that such frameworks are capable of achieving a peak test Sharpe Ratio of 5.65, a metric indicative of risk-adjusted return significantly exceeding traditional benchmarks. This level of performance isn’t simply about identifying profitable signals; it requires a system that simultaneously optimizes what is measured and how that measurement informs trading decisions, creating a self-improving cycle that maximizes profitability while minimizing exposure to market volatility. The resulting systems represent a paradigm shift in alpha generation, moving beyond prediction to a form of algorithmic adaptation.

Hyperparameter optimization with Optuna reveals that the evolved forecaster consistently achieves higher validation impact-adjusted PnL compared to the baseline forecaster across 120 trials.

MadEvolve: An LLM-Driven Evolutionary Framework

MadEvolve employs an iterative process where large language models (LLMs) generate novel trading strategies based on defined parameters and historical market data. This LLM-driven approach differs from traditional algorithmic strategy optimization by utilizing the LLM’s capacity for complex pattern recognition and creative combination of trading rules. The framework moves beyond simple parameter adjustments; it explores entirely new strategy concepts by prompting the LLM to propose and refine trading logic. Generated strategies are then evaluated against historical data to assess performance, with successful strategies forming the basis for further LLM-guided evolution and refinement. This process facilitates exploration of a broader strategy space than conventional methods, potentially identifying high-performing strategies previously unconsidered.

The MadEvolve framework utilizes an ‘LLM Ensemble’ comprised of multiple large language models operating in parallel to generate a wide array of candidate trading strategies. This ensemble approach moves beyond the limitations of a single LLM, allowing for exploration of a significantly broader solution space. Each LLM within the ensemble is prompted to create strategies based on varied inputs and constraints, resulting in a diverse population of potential algorithms. The combined output of this ensemble is then subjected to rigorous backtesting and evaluation, facilitating the identification of high-performing strategies that might not have emerged from a single LLM’s output. This method ensures a more comprehensive and robust optimization process.

Population management within the MadEvolve framework is crucial for sustaining both the diversity and performance of evolving trading strategies. This involves techniques to avoid premature convergence – where the population becomes overly homogeneous and exploration of the strategy space ceases. Specifically, strategies are evaluated based on performance metrics, with lower-performing individuals selectively removed and replaced by newly generated candidates from the LLM Ensemble. This continuous cycle of evaluation and replacement, coupled with mechanisms to encourage diversity, results in consistent Sharpe Ratio improvements ranging from 0.6 to 1.8 points when compared to baseline, non-evolved strategies. The system prioritizes maintaining a balance between exploitation of promising strategies and exploration of novel approaches, thereby maximizing the potential for long-term optimization.

In Run 1, the evolved strategy consistently outperformed the baseline, demonstrating positive cumulative impact-adjusted profit and loss across both the 2024 validation and 2025 test sets.

Rigorous Backtesting and Performance Evaluation

Backtesting is a fundamental component of evaluating the performance of trading strategies generated by MadEvolve, utilizing historical market data to simulate trade execution. This process allows for the assessment of a strategy’s potential profitability and risk characteristics before deployment in live markets. By applying the strategy to past data, researchers and developers can observe its behavior across different market conditions and identify potential weaknesses or areas for improvement. The quality of backtesting relies heavily on the accuracy and completeness of the historical data used, as well as the realism of the simulated trading environment. A robust backtesting process is essential for mitigating risk and increasing confidence in the efficacy of MadEvolve-generated strategies.

The trading simulation environment used for backtesting is designed to replicate real-world market conditions with a high degree of fidelity. This includes modeling order book dynamics, bid-ask spreads, and transaction costs. Specifically, the simulation accounts for latency, slippage, and the impact of order size on price movement. Data inputs incorporate historical order book data and volume to accurately represent liquidity and price formation. The simulation also models various market events, such as news releases and unexpected trading volume, to assess strategy robustness under diverse conditions. Furthermore, the environment allows for the adjustment of parameters related to exchange fees, regulatory constraints, and counterparty risk, providing a comprehensive assessment of potential impacts on strategy performance.

Performance evaluation of MadEvolve strategies utilizes Profit and Loss (Pnl) calculation alongside market impact assessment to determine both profitability and operational efficiency. Specifically, the evolved forecasting features demonstrate an R-squared value of 0.0105 when tested on the validation dataset, representing an improvement from its initial value of 0.0021. This metric indicates the proportion of variance in the dependent variable that is predictable from the independent variables within the forecasting model; a higher R-squared suggests a better fit of the model to the observed data, though the absolute value remains low indicating limited predictive power.

In Run 2, the evolved order placement strategy demonstrably outperformed the baseline, achieving higher cumulative impact-adjusted profit and loss on both the validation set (2024) and the test set (2025).

Order Execution and Strategic Implementation: Bridging the Gap

The efficacy of even the most meticulously crafted trading strategy hinges fundamentally on its successful execution as an order in the market. A profitable theoretical edge remains unrealized without a robust order execution process that efficiently converts signals into trades. This process isn’t merely about placing an order; it encompasses navigating market microstructure, minimizing transaction costs – including slippage and market impact – and maximizing the probability of achieving a desired fill price. Consequently, sophisticated traders and institutions dedicate substantial resources to optimizing order execution, recognizing it as the critical bridge between strategic insight and tangible financial gains. Without careful consideration of these factors, even a highly accurate strategy can be eroded by the realities of market dynamics, ultimately diminishing returns and jeopardizing profitability.

Passive limit-order execution represents a nuanced approach to translating trading signals into actualized positions, prioritizing minimal disruption to prevailing market conditions. Rather than aggressively attempting to immediately fill an order at the best available price – which can drive prices unfavorably – this strategy deploys orders as ‘limit orders’ placed slightly above the ask or below the bid. This method seeks to be filled only at a desired price or better, effectively allowing the market to come to the order. The benefit lies in reduced market impact – the degree to which a trade moves the price – and consequently, improved fill rates, as the order is less likely to trigger adverse price movements that would prevent its completion. By patiently awaiting favorable price levels, passive execution aims to secure better overall outcomes, especially for larger orders where immediate execution could significantly affect the market.

The synergy between MadEvolve and advanced execution platforms yields a remarkably robust and responsive trading solution. Rigorous statistical analysis confirms this performance isn’t simply attributable to random chance; validation tests exceeded 103.2 standard deviations, and subsequent testing reached 44.9 standard deviations. These substantial figures indicate a highly significant outcome, suggesting the integrated system consistently delivers results beyond the scope of typical market fluctuations. This adaptive capability allows for dynamic adjustments to trading parameters, optimizing execution in response to evolving market conditions and maximizing potential profitability. The platform’s ability to consistently outperform expectations is a testament to its sophisticated design and the efficacy of its underlying algorithms.

Enhancing Forecasts with Exponential Moving Averages: A Refined Signal

MadEvolve’s ‘Alpha Forecasting’ process gains increased sensitivity to recent price movements through the integration of Exponential Moving Averages (EMAs). Unlike simple moving averages which treat all past data equally, EMAs assign exponentially decreasing weights to older data, thereby emphasizing the most current information. This allows the system to react more swiftly to emerging short-term trends, potentially identifying opportunities and mitigating risks before they are fully reflected in longer-term indicators. By dynamically adjusting to these shifts, the forecasting engine demonstrates improved accuracy in volatile markets and provides a more responsive predictive capability. The implementation focuses on capturing transient patterns, enhancing the system’s ability to anticipate immediate market behavior.

The synergy between large language model (LLM)-driven evolution and dynamic technical indicators, such as exponential moving averages, culminates in a forecasting engine uniquely equipped to navigate volatile conditions. MadEvolve doesn’t rely on static models; instead, the LLM continuously refines the forecasting process, adapting to new data and market behaviors. This evolutionary approach, combined with the responsiveness of technical indicators to price movements, allows the engine to discern subtle trends and react with agility. The result is a system that not only predicts potential future outcomes but also strengthens its predictive capabilities over time, offering a robust and adaptive solution for forecasting in complex environments.

The evolution of MadEvolve’s forecasting engine doesn’t cease with the integration of Exponential Moving Averages; ongoing research actively seeks to elevate its predictive power through multifaceted enhancements. Investigations are currently underway to incorporate a wider array of data streams, moving beyond purely historical price action to include sentiment analysis from news sources, macroeconomic indicators, and even alternative data like social media trends. Furthermore, the team is exploring the potential of more sophisticated technical indicators, alongside machine learning algorithms capable of dynamically weighting these signals based on prevailing market conditions. This iterative process of refinement aims to create a truly adaptive forecasting system, capable of not just reacting to change, but anticipating it with increasing accuracy and robustness.

Analysis using PnL decomposition and scale-invariant ratios (Sharpe and Calmar) demonstrates that algorithmic improvements extend beyond simply scaling trade sizes, as evidenced by PnL ratios exceeding <span class="katex-eq" data-katex-display="false">1.0\times</span> and consistent performance improvements regardless of scale. — Analysis using PnL decomposition and scale-invariant ratios (Sharpe and Calmar) demonstrates that algorithmic improvements extend beyond simply scaling trade sizes, as evidenced by PnL ratios exceeding $1.0\times$ and consistent performance improvements regardless of scale.

The pursuit of robust trading strategies, as detailed in this work, echoes a fundamental need for communicative rationality. Jürgen Habermas observed, “The lifeworld is the horizon of meaning which is taken for granted.” This resonates with the LLM-driven evolution presented; the model doesn’t simply generate strategies, it refines them through iterative backtesting, effectively establishing a consensus – a shared understanding of what performs well within the historical data. This process mirrors the striving for intersubjective agreement, filtering noise and establishing strategies that aren’t merely statistical anomalies, but possess a degree of generalized validity. Beauty scales-clutter doesn’t-and the streamlined, performant systems resulting from this research exemplify that principle.

The Road Ahead

The demonstrated capacity of LLM-driven evolution to sculpt trading systems is not merely a quantitative improvement; it hints at a fundamental shift in how such systems are conceived. Yet, the elegance of a profitable strategy should not obscure the limitations. Current approaches, while mitigating p-hacking through rigorous backtesting, still grapple with the phantom of out-of-sample generalization. The market, after all, is not a static dataset, but a restless entity constantly reshaping its own contours.

Future work must venture beyond the pursuit of ever-smaller error metrics. A compelling direction lies in embedding more sophisticated models of market impact directly into the evolutionary process. A system that anticipates its own influence, that understands how its actions alter the very landscape it navigates, would represent a genuine leap forward. The current methods, for all their refinement, remain largely reactive; a truly intelligent system should, ideally, be anticipatory.

Ultimately, the quest is not simply for profitability, but for understanding. Each successful strategy, each optimized parameter, should whisper insights into the underlying dynamics of the market. If the tools remain opaque, if the evolved systems are black boxes, then the endeavor, however lucrative, remains incomplete. The most elegant solutions are not those that merely work, but those that reveal why they work.

Original article: https://arxiv.org/pdf/2605.23007.pdf

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