The Gaining Paths to Investment Success: Information-Driven LLM Graph Reasoning for Venture Capital Prediction

In one sentence

Summary

The paper addresses the challenge of predicting early-stage startup success in venture capital, a task requiring synthesis of complex relational evidence that traditional machine learning and graph neural networks cannot interpretably reason over. While LLMs offer reasoning capabilities, they face a modality mismatch with graph structures, and existing graph-LLM methods focus on in-graph tasks rather than off-graph prediction where the target outcome lies outside the network. The core problem is selecting high-value graph paths that maximize predictor performance on an external objective while enabling step-by-step, interpretable reasoning for investment theses. The proposed MIRAGE-VC framework employs an information-gain-driven path retriever that iteratively selects neighbors to maximize LLM predictor accuracy, distilling investment networks into compact chains to avoid context overflow. It integrates three evidence streams—company disclosures, lead investor profiles, and graph-based paths—using a multi-agent architecture where specialist agents analyze each stream. A learnable gating network dynamically weights these heterogeneous evidence sources based on company attributes, and a manager agent synthesizes the weighted signals into a calibrated prediction with an interpretable rationale. Experiments use the PitchBook dataset with strict anti-leakage controls, evaluating on startups that secured Series A funding within one year of their seed round. MIRAGE-VC achieves state-of-the-art results, improving F1 by 5.0% and Precision@5 by 16.6% over strong baselines like GNN-RAG and SSFF. The model demonstrates that correctly classified startups are associated with longer, richer retrieved paths, highlighting the value of structural context. The approach provides actionable investment signals and interpretable rationales, addressing the lack of transparency in black-box GNNs. Limitations include reliance on the proprietary PitchBook dataset, which restricts public reproducibility, and a myopic supervision objective that optimizes local one-hop gains rather than globally optimal subgraphs, potentially overlooking long-term dependencies in the network.

Research problem

Predicting early-stage startup success in venture capital requires synthesizing complex relational evidence from company disclosures, investor track records, and investment network structures. Traditional ML and GNNs lack interpretability and reasoning capabilities, while LLMs struggle with graph modality and context limits. The specific challenge is selecting graph paths that maximize performance on an off-graph prediction task (startup success) while enabling explicit, step-by-step reasoning.

Core contributions

• Proposes MIRAGE-VC, a multi-perspective RAG framework for VC prediction that addresses path explosion and heterogeneous evidence fusion.
• Introduces an information-gain-driven path retriever that iteratively selects high-value neighbors to distill investment networks into compact, reasoning-ready chains.
• Develops a multi-agent architecture with a learnable gating mechanism to dynamically weight company, investor, and graph evidence based on startup attributes.
• Demonstrates state-of-the-art performance on real-world VC data with strict anti-leakage controls, improving F1 by 5.0% and Precision@5 by 16.6%.
• Provides interpretable, chain-of-thought investment theses, bridging the gap between black-box predictions and human-readable due diligence.

Method

• Information-gain-driven path retriever: Uses a frozen LLM to compute marginal gain of adding nodes to a path, training a lightweight MLP selector via listwise loss.
• Text retrieval: Retrieves semantically similar peer companies and lead investor profiles using frozen sentence encoders, with temporal leakage prevention.
• Multi-agent analysis: Three specialist agents (Peer-Company, Investor Profile, Investment Chain) analyze their respective evidence streams using frozen GPT-3.5 Turbo.
• Learnable gating network: Embeds agent rationales and company attributes to learn instance-specific weights for adaptive evidence fusion.
• Manager agent: Synthesizes weighted agent outputs into a final binary prediction and natural language rationale.
• Evaluation metrics: AP@K, Precision, Recall, F1, with monthly-averaged AP@K to assess temporal consistency.

Results & findings

• MIRAGE-VC achieves +5.0% F1 and +16.6% Precision@5 over the strongest baselines.
• AP@5 improves by 16.6%, AP@10 by 16.7%, and AP@20 by 8.0%, showing superior top-k recommendation capability.
• Ablation studies confirm the necessity of graph retrieval, path selection, and adaptive gating for performance.
• Correctly classified startups are associated with longer retrieved paths (4.44 vs 3.31 hops), indicating richer context aids reasoning.
• The model outperforms GNNs in precision and LLM-only methods in structural reasoning, balancing recall and precision.
• Results are consistent across different backbone LLMs, with GPT-4o-mini showing further relative improvements.

Data & experiments

• Uses the PitchBook Global VC dataset (2005-2023) containing 263,729 startups and 1,014,157 individuals.
• Defines success as securing Series A funding within 12 months of the first formal financing round (seed/angel).
• Evaluates on 2,510 startups funded between Oct 2021 and Nov 2023 to prevent data leakage from LLM pretraining.
• Compares against GNN-based (SHGMNN, GST), embedding-based (BERT Fusion), RAG-based (Standard RAG, GNN-RAG), and LLM-driven (SSFF) baselines.
• Metrics include AP@K (Average Precision at K), Precision, Recall, and F1 score, with AP@K highlighting top-k recommendation quality.

Limitations & caveats

• Relies exclusively on the proprietary PitchBook dataset, limiting public reproducibility and generalizability to other data sources.
• The supervision objective is myopic, optimizing local one-hop gains rather than globally optimal subgraphs, potentially missing long-term dependencies.
• Evaluation is limited to predicting Series A funding within 12 months, which may not capture broader long-term startup success metrics.
• The framework's performance depends on the quality of the underlying LLM and the completeness of the investment network data.
• The gating network learns weights based on historical data, which may not adapt well to novel startup types or market shifts without retraining.

Curator notes

Discovery status: institutional_trading_focus_promoted. Source links: LLM venture capital due diligence investment