An agent-based framework for systematic generative feature engineering in financial markets

Abstract

It runs in an iterative process over 30 loops, generating a candidate feature set at each iteration. This iterative approach allows the performance feedback of earlier loops to inform the creative direction of later ones. To ensure that the generated feature pool does not include duplicates, I employ a four-layer deduplication mechanism at increasingly deeper levels: exact-name matching; hash-based syntactic camouflage detection; LLM-powered mathematical uniqueness verification; and semantic similarity analysis in the expression space. Every generation prompt includes randomly chosen mathematical concepts sampled from a broad space of subject areas, including fractal geometry, information theory, and spectral analysis, so that the system does not hone in on familiar feature archetypes over successive iterations. This gradation of difficulty also applies to the sampling temperature and novelty requirements, meaning that the system is forced to explore genuinely new mathematical territory throughout later iterations. As the Feature Generator, five large language model configurations are evaluated under similar experimental conditions, testing with ChatGPT 4o (with and without its mini version), ChatGPT 5, ChatGPT 5.2, and Claude Opus 4.6. The cross-source deduplication step selects the features with the largest scores from each configuration, which are then combined with the Case 0 baseline set to provide results for comparative evaluation. All features are evaluated in the same environment to ensure that any discrepancies between types of features observed come from the quality of generated features and not variance in methodology. These configurations all improve upon the classification accuracy and simulated model return achieved by Case 0, confirming the main hypothesis of this thesis. ChatGPT 4o outperforms in the majority of evaluation metrics, producing features that yield the best mean accuracy and highest economic returns.The single best feature it produces reflects accuracy and return figures that far exceed those in the traditional library. ChatGPT 4o mini is hot on its heels, delivering the best directional profile across all configurations. The other configurations show gradual, but still positive, improvements over the baseline. Interestingly, ChatGPT 4o beats both of the latest GPT models, suggesting that creative diversity in the context of generation is more important than model size for this task. Although the magnitude varies, there is a systematic long bias across all configurations. Claude Opus 4.6 stands out for a strong directional bias toward upward predictions and weak performance in providing a reliable signal for short positions. Conversely, ChatGPT 4o mini provides the most balanced predictions and demonstrates the most powerful short-side performance across all LLM configurations. Meanwhile, momentum features are continually hailed as the best overall performers across the top-ranked models, corroborating the well-established fact that opinion is expressed in monetary investments through long or short signals, and that such a generative system overall provides such evidence without explicit domain guidance. Additionally, different models exhibit distinct preferences for categories, indicating substantial disparities in their use of training data. The main caveats of the study are that it examines large-cap U.S. equities in a generally bull market, has not accounted for transaction costs in the performance numbers, and that LLM reasoning is opaque, making formal attribution of predictive power to particular mathematical mechanisms difficult.