Simulating investors as interacting agents and learning over co-investment networks materially improves prediction of which startups secure financing; on PitchBook data the model raises average precision@10 by about 25% versus standard baselines. While providing multiperspective, interpretable reasoning about group investment choices, the approach rests on simulated investor traits and a single data source, limiting immediate generalization.

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

Liu, Zhongyang, Pei, Haoyu, Xiao, Xiangyi, Du, Xiaocong, Li, Yihui, Hong, Suting, Zhang, Kunpeng, Zhang, Haipeng · December 27, 2025 · arXiv (Cornell University)

openalex correlational medium evidence 7/10 relevance Full text usable extracted full text Source PDF

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

OpenAlex

Latest observation: August 25, 2026

Liu, Zhongyang provider ID
Pei, Haoyu provider ID
Xiao, Xiangyi provider ID
Du, Xiaocong provider ID
Li, Yihui provider ID
Hong, Suting provider ID
Zhang, Kunpeng provider ID
Zhang, Haipeng provider ID

Semantic Scholar

Latest observation: August 25, 2026

Zhongyang Liu provider ID
Haoyu Pei provider ID
Xiangyi Xiao provider ID
Xiaocong Du provider ID
Yihui Li provider ID
Suting Hong provider ID
Kunpeng Zhang provider ID
Haipeng Zhang provider ID

Framing startup financing as a multi-agent group decision process and using a GNN over a co-investment network improves prediction of financing outcomes on PitchBook data, with roughly 25% relative gain in average precision@10 compared to baselines.

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

0 cumulative citations

OpenAlex · Observed August 25, 2026

View corpus context

1 cumulative citations

Semantic Scholar · Observed August 25, 2026

View corpus context

Due to the high value and high failure rate of startups, predicting their success has become a critical challenge across interdisciplinary research. Existing approaches typically model success prediction from the perspective of a single decision-maker, overlooking the collective dynamics of investor groups that dominate real-world venture capital (VC) decisions. In this paper, we propose SimVC-CAS, a novel collective agent system that simulates VC decision-making as a multi-agent interaction process. By designing role-playing agents and a GNN-based supervised interaction module, we reformulate startup financing prediction as a group decision-making task, capturing both enterprise fundamentals and the behavioral dynamics of potential investor networks. Each agent embodies an investor with unique traits and preferences, enabling heterogeneous evaluation and realistic information exchange through a graph-structured co-investment network. Using real-world data from PitchBook and under strict data leakage controls, we show that SimVC-CAS significantly improves predictive accuracy while providing interpretable, multiperspective reasoning, for example, approximately 25% relative improvement with respect to average precision@10. SimVC-CAS also sheds light on other complex group decision scenarios.

Summary

Main Finding

SimVC-CAS — a roleplay-based collective agent system that models venture-capital (VC) decisions as multi-agent interactions over a co-investment graph — materially improves early-stage startup financing prediction versus single-decision-maker baselines. On proprietary and public VC data (with strict anti-leakage controls) the system achieves roughly a 25% relative improvement in average precision@10; its interaction module (VGAT) also yields substantially better link-prediction performance (F1 79.31% vs 73.47% for a GAT baseline). The model is interpretable at the agent-reasoning level and is especially effective for startups that are network-central.

Key Points

Motivation
- Standard automated methods treat financing prediction as a single-investor decision problem, ignoring syndicated/co-investor dynamics that shape real outcomes.
- Real VC outcomes emerge from heterogeneous investor judgments and networked peer influence; average graph distance among co-investors in the data is small (~2.16), indicating strong local interaction effects.
Conceptual innovation
- Reformulates financing prediction as a collective-group decision task: a pool of investor agents evaluate a startup, interact over a co-investment network, revise beliefs, and produce an aggregate investment confidence score.
- Leverages LLM role-playing to encode heterogeneous investor personas (preferences, biases, histories) for realistic, diverse initial judgments.
Architecture components
Startup Panoramic Portrait: comprehensive startup context (basic info, team, financing history, affiliated companies).
Heterogeneous Investor Portraits: per-agent profiles (personal, employment, early-/other-stage investment records); candidate investors sampled from historical investors and their co-investment partners.
Collective Interaction Modeling: Virtual-Node Graph Attention Network (VGAT) — the startup is a virtual node connected to all investors; uses global GAT, local GAT, and a cross-attention mechanism (virtual node as query over edge representations) to predict effective interactions (proxied by future co-investment).
Learning signals and decision flow
- VGAT trained with cross-entropy (future co-investment labels) plus a contrastive loss on edge embeddings.
- Investor agents produce initial decisions via a frozen LLM conditioned on persona + startup; interaction edges inferred by VGAT define neighbors whose initial decisions are shared; agents then update decisions via the LLM conditioned on neighbors’ views; final investment confidence = fraction of agents choosing to invest.
Empirical highlights
- VGAT link prediction: Precision 76.18 / Recall 82.70 / F1 79.31 vs GAT 69.62 / 77.76 / 73.47.
- End-to-end financing prediction: ~25% relative improvement in average precision@10 compared to single-perspective baselines (paper reports this as the primary aggregate gain).
- Interaction effects are strongest for startups with high network centrality; agent-level reasoning analysis provides interpretable traces of how network signals change beliefs.

Data & Methods

Datasets
- PitchBook global VC dataset (2005–Nov 2023): ~263,729 startups and ~1,014,157 individuals. Main evaluation subset: 2,507 startups whose initial funding occurred Sep 2021–Nov 2022; outcome = whether a startup obtains follow-on financing within one year (533 positives, 1,974 negatives).
- Authors report experiments on both proprietary and public VC datasets with temporal split and anti-leakage procedures; VGAT was trained on data before Sep 2021 and tested on later data.
Feature construction
- Startup features: founding date, industry, description, products/keywords, team member backgrounds, financing history, affiliated companies.
- Investor features: name/demographics, employment history, detailed past investments (split into early-stage vs other), structured summaries/embeddings of invested startups.
Modeling details
- Role-playing agents: frozen LLMs produce initial binary invest / not-invest outputs given (startup, investor profile).
- Node embeddings: encoder (authors used jina-colbert-v2 for long-context embedding) to embed startup and investor profiles for graph modeling.
- VGAT: graph = investor nodes VR plus virtual node v_target; edges = historical co-investment links (VR×VR) and virtual edges (v_target to all VR). Processing: global GAT → local GAT on VR → virtual node MLP query → cross-attention over edge representations to compute weighted edge embeddings. Edge prediction is framed as binary classification (future co-invest vs not).
- Loss: L = CrossEntropy(edge labels) + λ * ContrastiveLoss(edge embeddings).
- Decision update: agents receive neighbors’ initial decisions and neighbor profiles, then LLM produces revised decision; aggregate score = proportion investing.
Evaluation
- Link prediction performance for VGAT vs GAT reported (precision/recall/F1).
- End-task startup financing prediction evaluated with ranking/precision metrics (average precision@10 emphasized).
- Temporal train/validation/test splits used to avoid leakage.
Code
- Codebase made available by authors (link referenced in paper).

Implications for AI Economics

Modeling market microstructure of VC decisions
- SimVC-CAS provides a formal, data-driven way to model how investor heterogeneity + network topology jointly determine funding outcomes. This enables deeper study of how local peer influence and co-investor composition produce distributional outcomes (who gets funded and why).
Forecasting and resource allocation
- Improved predictive performance for early-stage financing can enhance automated deal sourcing, benchmarking, and allocation tools for limited partners, accelerators, and platforms — but also risks entrenching algorithmic selection biases if used uncritically.
Policy and market-design use cases
- The framework can be used to simulate counterfactual interventions: e.g., adding or removing investor types, changing information flows, or simulating policy nudges (matching grants, disclosure rules) to analyze their effects on financing inequality, market concentration, or systemic fragility.
Research into inequality, network externalities, and systemic risk
- By quantifying how network centrality amplifies interaction effects, the method supports empirical work on preferential attachment, rich-get-richer dynamics, and path dependence in entrepreneurial finance.
Interpretability and causal investigation
- Agent-level roleplay traces offer interpretable signals about why an investor’s view changed after peer influence — useful for behavioral analyses of biases, herding, and informational cascades. However, causal claims require careful design (the simulation approximates but does not prove real-world causality).
Caveats & ethical considerations
- Dependence on LLM role-playing can import model biases and fictive reasoning patterns; future-co-investment as a proxy for real interaction is imperfect and may conflate correlated preferences with direct communication.
- Proprietary data access and privacy: real investor communications are private; model users must respect confidentiality and be careful about deploying predictive tools that affect real entrepreneurs’ access to capital.
- Potential for reinforcing selection dynamics: automated systems informed by this approach could unintentionally amplify network-based exclusion or concentration unless countermeasures are implemented.

Limitations noted by the authors and to consider for follow-up work: reliance on LLM-generated personas and frozen LLMs, proxy-labeling of interaction via future co-investment, dataset biases (geography/sector), computational cost, and the need for randomized or quasi-experimental validation to support causal claims about network interventions.

Assessment

Paper Typecorrelational Evidence Strengthmedium — The paper demonstrates meaningful predictive improvements (≈25% relative gain on average precision@10) using real-world PitchBook data and reports strict data-leakage controls, which supports the claim that the model predicts financing outcomes better than baselines; however, the evidence is limited to predictive performance on a single commercial dataset without external validation, robustness checks across markets/stages, or causal identification of mechanisms. Methods Rigormedium — Methodological strengths include modeling group decision dynamics with heterogeneous role-playing agents, using a graph-structured co-investment network and a GNN-based interaction module, and attention to data-leakage controls; methodological weaknesses are potential sensitivity to agent design choices and hyperparameters, limited description of validation (e.g., temporal/holdout splits, sectoral robustness), and reliance on supervised prediction rather than causal inference or randomized evaluation. SampleProprietary real-world data from PitchBook comprising startups, firm-level fundamentals, funding outcomes, investor identities and co-investment relations; used to construct a co-investment graph and train a supervised GNN + multi-agent simulation to predict financing/success outcomes (paper does not specify full sample size or time span in the summary). Themesinnovation human_ai_collab GeneralizabilitySingle commercial data source (PitchBook) may be biased toward certain geographies, sectors, and observable investors, Model performance may not generalize across stages (seed vs. late-stage), industries, or time periods without additional validation, Role-playing agent specifications and assumed investor traits may not reflect real-world heterogeneity and strategic behavior, High data and modeling complexity (GNN + multi-agent design) limits applicability where co-investment networks or detailed investor features are unavailable, Predictive improvement does not establish causal relationships about what drives startup success

Claims (7)

Claim	Direction	Outcome	Confidence & Evidence	Details
We propose SimVC-CAS, a novel collective agent system that simulates VC decision-making as a multi-agent interaction process. Other	positive	ability to simulate VC decision-making	Reading fidelity high Study strength high	not reported 0.5
The model uses role-playing agents and a GNN-based supervised interaction module. Other	positive	model architecture components (use of role-playing agents and GNN interaction module)	Reading fidelity high Study strength high	not reported 0.5
We reformulate startup financing prediction as a group decision-making task that captures both enterprise fundamentals and the behavioral dynamics of potential investor networks. Decision Quality	positive	quality of group decision-making formulation for startup financing prediction	Reading fidelity high Study strength medium	not reported 0.3
Each agent embodies an investor with unique traits and preferences, enabling heterogeneous evaluation and realistic information exchange through a graph-structured co-investment network. Decision Quality	positive	heterogeneous investor evaluation and information exchange realism	Reading fidelity high Study strength high	not reported 0.5
We use real-world data from PitchBook and enforce strict data leakage controls in experiments. Other	positive	use of real-world PitchBook dataset and application of data leakage controls	Reading fidelity high Study strength medium	not reported 0.3
SimVC-CAS significantly improves predictive accuracy, e.g., approximately 25% relative improvement with respect to average precision@10. Decision Quality	positive	average precision@10 (predictive accuracy)	Reading fidelity high Study strength medium	approximately 25% relative improvement with respect to average precision@10 0.3
SimVC-CAS provides interpretable, multiperspective reasoning and can shed light on other complex group decision scenarios. Decision Quality	positive	interpretability and generalizability to other group decision scenarios	Reading fidelity medium Study strength low	not reported 0.09