Master Modeler: Learning Institutional Portfolio Construction with Large Language Models

Abstract

Investment institutions exert substantial influence on asset prices, liquidity, and market stability, making the ability to forecast their portfolio adjustments both academically and practically important. We study modeling investment institutions by fine-tuning large language models (LLMs) to predict next-quarter changes in holdings from publicly disclosed information under partial observability. Using SEC EDGAR as a unified data source, we construct a controlled experimental setting that combines structured financial fundamentals with optional document-level signals from MD&A disclosures, and explicitly contrast three paradigms: sequence models trained on tabular fundamentals, LLMs operating solely on serialized numerical inputs, and document-aware LLMs augmented with semantic summaries of MD&A narratives. We cast holdings-change prediction as a constrained structured generation task, jointly modeling discrete adjustment decisions and numerical outcomes with a fixed output schema, and adopt response-only supervision with imbalance-aware training to improve interpretability and robustness. Experiments on U.S. institutional holdings from 2015 to 2024 demonstrate consistent gains: language-based sequence modeling outperforms recurrent numerical baselines even without textual inputs, and incorporating MD&A-based summaries yields further improvements, confirming that disclosure narratives provide complementary predictive value beyond structured fundamentals alone. The results suggest that institutional portfolio adjustments are not fully captured by numerical dynamics in isolation and that document-level semantic understanding is critical for effective institution modeling.