Beyond Isolated Investor: Predicting Startup Success via Roleplay-Based Collective Agents

In one sentence

Summary

The paper addresses the limitation of existing startup success prediction models that treat investors as isolated decision-makers, ignoring the collective dynamics and network effects inherent in real-world venture capital syndication. It reframes financing prediction as a group decision-making task, positing that investor interactions and peer influence significantly impact funding outcomes. The proposed SimVC-CAS system simulates this process by constructing heterogeneous investor agents and modeling their interactions within a graph-structured co-investment network. The method employs a three-module architecture: a startup panoramic portrait integrating firm fundamentals and team data; heterogeneous investor portraits generated from historical investor profiles and past investment behaviors; and a collective interaction module using a Virtual Node Graph Attention Network (VGAT). VGAT models the startup as a virtual node to capture context-specific interaction patterns among investors. Agents perform initial individual evaluations using LLM role-playing, then update decisions based on peer influence derived from the learned interaction topology. The system is trained and evaluated on proprietary PitchBook data and public Crunchbase data, using strict temporal splits to prevent leakage. SimVC-CAS achieves a 25% relative improvement in average precision@10 compared to the strongest baseline, demonstrating superior capability in ranking high-potential startups. The interaction mechanism is particularly effective for network-central startups, confirming the importance of network structure. Analysis reveals that correct decision revisions are driven more by authority influence and information completion than consensus pressure. The system exhibits high consistency with real investor decisions and interpretability, though it relies on proprietary data and may face generalization challenges to other group decision contexts.

Research problem

Predicting early-stage startup success (specifically securing follow-on financing) is critical but challenging due to high failure rates. Existing automated methods typically model this from a single decision-maker's perspective, failing to capture the collective dynamics, heterogeneous preferences, and network effects that dominate real-world venture capital decision-making. The paper aims to bridge this gap by modeling VC decisions as a multi-agent interaction process.

Core contributions

• Introduces SimVC-CAS, a collective agent system that reframes startup financing prediction as a group decision-making task.
• Proposes a novel interaction mechanism using VGAT to model investor dynamics and startup-centric interaction patterns.
• Demonstrates significant performance improvements (25% relative gain in AP@10) over single-investor and non-interactive multi-agent baselines.
• Provides interpretability by analyzing agent reasoning for decision changes, linking network environment to decision quality.
• Validates the model's consistency with real investor decisions and its effectiveness for network-central startups.

Method

• Startup Panoramic Portrait: Integrates basic info, team composition, financing history, and affiliated companies.
• Heterogeneous Investor Portraits: Profiles based on personal info, past investment experience, and employment history.
• LLM Role-Playing: Agents simulate individual investor evaluations using frozen LLMs conditioned on profiles.
• VGAT Interaction: Uses a virtual node (startup) to attend to investor nodes and edges in a co-investment network.
• Cross-Attention Mechanism: Computes attention scores for co-investment edges based on startup context.
• Decision Update: Agents revise initial decisions based on peer influence from interaction neighbors.
• Confidence Score: Aggregates individual investment decisions to rank startups.
• Loss Function: Joint optimization using Cross-Entropy and Contrastive Loss for link prediction.

Results & findings

• SimVC-CAS achieves 37.52% AP@10, a 25% relative improvement over the strongest baseline (SSFF at 30.02%).
• The model outperforms GNN-based methods (SHGMNN, GST) and LLM-RAG variants (GNN-RAG).
• Decision Consistency with real investors is 0.71, significantly higher than the random baseline of 0.50.
• Interaction mechanism is most effective for startups with high network centrality (PageRank).
• Correct decision revisions are driven by Authority Influence (24.7%) and Information Completion (11.3%), not just Consensus Pressure.
• Ablation shows that removing role-playing or interaction significantly degrades performance.
• Full interaction (all-to-all) performs worse than no interaction, highlighting the need for structured network-based interaction.

Data & experiments

• Proprietary PitchBook dataset (2005-2023) with 263,729 startups and 1,014,157 individuals.
• Public Crunchbase snapshot (2013) for cross-dataset validation.
• Task: Predicting whether a startup secures follow-on funding within one year of initial financing.
• Metrics: Average Precision@K (AP@K), Precision, Recall, F1, and Decision Consistency.
• Baselines: Random, BERT Fusion, SHGMNN, GST, GNN-RAG, SSFF.
• Strict temporal split to prevent data leakage; GPT-3.5 used for inference on post-cutoff data.

Limitations & caveats

• Relies on proprietary PitchBook data, limiting reproducibility and generalizability.
• Public Crunchbase data is outdated (2013), potentially affecting cross-dataset validation.
• Data leakage risk from newer LLMs (GPT-4o, Claude 3.5) used in ablation studies.
• Decision Consistency metric may be biased by the rarity of 'Invest' decisions in raw data.
• Model assumes historical co-investment relationships are valid proxies for interaction.
• Computational cost of simulating multiple agents and interactions may be high.

Curator notes

Discovery status: institutional_trading_round2_promoted. Source links: The Gaining Paths to Investment Success: Information-Driven LLM Graph Reasoning for Venture Capital Prediction