Can AI Play Venture Capitalist?

Author: Denis Avetisyan

A new framework simulates the dynamics of venture capital investment using interacting AI agents to predict startup success with greater accuracy.

Unlike prior approaches focused on isolated decision-making, this work models the interplay between startups and investors by simultaneously considering the viewpoints of all involved parties, offering a more holistic understanding of collective behavior.

SimVC-CAS leverages multi-agent systems, graph neural networks, and large language models to model co-investment behavior and improve investment prediction.

Predicting startup success remains a critical yet challenging task given the high-risk, high-reward nature of venture capital. This paper, ‘LLM Agents as VC investors: Predicting Startup Success via RolePlay-Based Collective Simulation’, introduces SimVC-CAS, a novel framework that models VC decision-making as a multi-agent interaction process, significantly improving prediction accuracy by simulating realistic investor dynamics. By framing investment prediction as a group decision-making task-capturing both fundamental enterprise data and behavioral nuances-SimVC-CAS achieves approximately 25% relative improvement in precision. Could this approach to collective intelligence unlock more interpretable and effective solutions for complex decision-making beyond the realm of venture capital?

Deconstructing the Oracle: Why Venture Capital Prediction Fails

The pursuit of forecasting startup viability is central to the practice of venture capital, yet traditional predictive models frequently fall short due to their inability to fully represent the intricate relationship between investment firms and the companies they fund. These models often treat startups as isolated entities, neglecting the significant impact of investor involvement – from strategic guidance and network access to subsequent funding rounds – on a company’s trajectory. This simplification overlooks the reciprocal influence where investor actions shape startup development, and startup performance, in turn, affects investor portfolios. Consequently, reliance on static analyses of limited data points, such as initial funding amounts or market size, can result in inaccurate assessments and a failure to identify genuinely promising ventures within a dynamic and interconnected ecosystem.

Conventional venture capital prediction models often fall short because they treat startup characteristics as fixed entities, neglecting the crucial, evolving relationship between investors and the ventures they fund. These static approaches typically focus on easily quantifiable features – like initial funding amounts or founder demographics – but fail to capture the dynamic interplay of information exchange, mentorship, and strategic pivots that define a startup’s trajectory. Furthermore, the significant influence of investor networks – the connections, shared knowledge, and signaling effects within venture capital communities – are largely ignored. A startup’s success isn’t solely determined by its initial attributes; it’s profoundly shaped by how investors perceive and react to its progress, and how those investors leverage their networks to support or abandon the venture, creating a feedback loop that static models cannot replicate.

The venture capital ecosystem’s inherent dynamism presents a significant challenge to predictive modeling. Traditional methods, often reliant on historical data and static firm or founder characteristics, struggle to account for the rapid shifts in technology, market trends, and investor sentiment. Consequently, predictions based on these approaches frequently diverge from actual outcomes, leading to suboptimal investment decisions and unrealized opportunities. This isn’t merely a matter of statistical error; the very landscape of innovation is constantly reshaping itself, rendering past performance an imperfect guide to future success. The inability to accurately forecast startup trajectories thus creates a cycle of missed potential, where promising ventures may be overlooked and capital misallocated, ultimately hindering growth and innovation within the broader economic system.

Investors iteratively refine their investment decisions by interacting with and observing the actions of other investors.

SimVC-CAS: Replicating the Ecosystem, Not Just the Data

SimVC-CAS is a computational system designed to replicate the dynamics of the venture capital (VC) ecosystem. It functions as a multi-agent system wherein individual agents represent both investors and startups, and their interactions are the core of the simulation. The system aims to model the complete VC decision-making process, from initial contact and due diligence to investment and follow-on funding rounds. By simulating these interactions at scale, SimVC-CAS enables researchers to study factors influencing investment outcomes, assess the impact of network effects, and explore counterfactual scenarios within the VC landscape. The system does not rely on pre-defined investment rules but instead allows agent behavior to emerge from simulated interactions and individual preferences.

The Co-Investment Network within SimVC-CAS is modeled using a Graph Attention Network (GAT) to explicitly represent and leverage relationships between investors. This network structures investors as nodes and co-investment occurrences as edges, allowing the model to quantify the influence each investor has on others during the decision-making process. The GAT architecture enables weighted aggregation of features from neighboring investors, meaning an investor’s decision is not solely based on startup characteristics, but also on the preferences and actions of their co-investors. Attention weights are learned during training, dynamically adjusting the importance of each relationship and capturing varying degrees of influence within the network. This approach allows SimVC-CAS to simulate how information and preferences propagate through the venture capital community, impacting investment decisions beyond simple individual assessments.

The SimVC-CAS system employs a Virtual Node to mediate interaction between investors and target startups, facilitating a more detailed representation of investor attention than direct connections would allow. This node functions as a central point of contact, receiving signals from investors regarding their evaluation of the startup and relaying that aggregated attention – quantified by investor weights and signals – to the startup. Critically, the Virtual Node does not represent a real-world entity, but a computational construct enabling the system to model the degree to which investors are focused on a specific startup, factoring in the strength of their relationships and individual preferences without requiring direct, pairwise connections between all actors in the co-investment network.

Investor agents within SimVC-CAS are instantiated using Large Language Model (LLM) role-playing to replicate individualized investment strategies. Each investor agent is provided with a defined persona encompassing investment focus, risk tolerance, and typical investment amounts, which are then used to prompt the LLM. This allows the agent to evaluate startup pitches and make investment decisions consistent with its assigned profile. The LLM generates responses – including investment amounts or rejections – based on the input pitch and the agent’s internal persona, effectively simulating heterogeneous investor behavior within the co-investment network. This approach moves beyond simple rule-based systems by leveraging the LLM’s capacity for nuanced judgment and complex reasoning.

Unveiling Predictive Power: Graph Modeling and Multi-Agent Interaction

The SimVC-CAS system employs collective agent interaction to simulate venture capital investment decisions, moving beyond predictive models reliant solely on startup feature analysis. This approach utilizes multiple independent agents, each representing an investor, that interact with decomposed startup data – specifically utilizing the SSFF framework – and with each other. By simulating these interactions, the system captures emergent behaviors and nuanced investment strategies that are not readily apparent through static feature evaluation. The collective output of these agents, reflecting their individual assessments and negotiated valuations, provides a more holistic and realistic evaluation of startup potential compared to single-agent or feature-based prediction methods.

SimVC-CAS employs a Structured Semantic Feature Framework (SSFF) to disaggregate raw startup data into independent, analyzable components. This decomposition process separates information regarding the founding team, the product or service, market dynamics, and financial projections into discrete feature sets. By providing each agent within the simulation with access to these isolated datasets, SSFF facilitates a more nuanced and comprehensive evaluation of startup potential, as agents can independently assess each component before contributing to a collective assessment. The framework supports a granular analysis, moving beyond holistic feature vectors to pinpoint specific strengths and weaknesses within the startup’s profile.

The SimVC-CAS system employs Vanishing Gradient Attention Networks (VGATs) as a core component for modeling investor interactions. VGATs, a type of graph neural network, address limitations in standard Graph Convolutional Networks (GCNs) by incorporating attention mechanisms that dynamically weight the importance of neighboring nodes during message passing. This allows the model to focus on the most relevant connections within the investor network, capturing complex relationships beyond simple adjacency. Specifically, VGATs utilize learnable attention coefficients to mitigate the vanishing gradient problem often encountered in deep graph neural networks, enabling effective propagation of information across multiple hops and improving the model’s ability to discern nuanced patterns in investor behavior and influence.

To improve the predictive capabilities of our Graph Neural Network (GNN) models, we integrated a Retrieval-Augmented Generation (RAG) framework. This involves retrieving relevant external knowledge – such as news articles, market reports, and industry analyses – based on the startup and investor data represented in the graph. This retrieved information is then incorporated as contextual input to the GNN, allowing the model to consider factors beyond the initial feature set and graph structure. Specifically, the RAG component provides the GNN with additional data points related to market trends, competitive landscapes, and investor preferences, thereby enhancing its ability to assess startup potential and predict investment outcomes. This process effectively addresses limitations inherent in relying solely on the information contained within the graph itself.

Beyond Prediction: Deciphering Investment Dynamics and Mitigating Bias

SimVC-CAS moves beyond simply forecasting which startups will succeed, offering a detailed understanding of why certain ventures are favored by investors. The framework doesn’t just output a prediction; it illuminates the key factors driving those assessments, revealing the relative importance of variables like team experience, market size, and funding history. This granular insight is achieved through a sophisticated analytical process, allowing stakeholders to dissect investment decisions and identify underlying patterns. By exposing these influential factors, SimVC-CAS provides a powerful tool for understanding investment dynamics and potentially mitigating biases that might otherwise hinder objective evaluation – ultimately fostering a more informed and equitable venture capital landscape.

The SimVC-CAS framework demonstrates a substantial advancement in startup success prediction, exceeding the performance of existing baseline methods. Specifically, the model achieves an Average Precision at K=10 (AP@10) of 37.52%, representing a significant 25.0% improvement over prior approaches. This heightened precision at the top of ranked predictions indicates a greater ability to identify truly promising startups early in the evaluation process. The model’s capacity to accurately prioritize potential investments has implications for venture capitalists and other stakeholders seeking to maximize returns and efficiently allocate capital, suggesting a powerful new tool for navigating the competitive landscape of startup funding.

The SimVC-CAS framework demonstrates substantial performance gains not only in identifying top startups, but also in ranking them effectively at scale. Evaluation metrics reveal an Average Precision at K=20 (AP@20) of 33.15%, representing an 18.8% improvement over existing methods. This means the model successfully places relevant startups within the top 20 recommendations almost a fifth more often. Further extending this capability, the framework achieves an Average Precision at K=30 (AP@30) of 30.14%, a 13.1% improvement, indicating continued precision even when considering a larger pool of potential investments. These results highlight the system’s ability to consistently identify and prioritize promising ventures, offering a significant advantage in navigating the complex landscape of startup funding.

The SimVC-CAS framework extends beyond simple prediction to model the nuances of investor behavior. Through the creation of diverse investor profiles – each with unique risk tolerances, investment preferences, and potential biases – the system simulates how these factors can skew capital allocation. This allows for a detailed analysis of how inherent biases, conscious or unconscious, might lead to underfunding of promising ventures led by underrepresented groups, or overinvestment in ventures mirroring the investor’s own background. By identifying these patterns in simulated outcomes, the framework facilitates the development of targeted strategies – such as blind reviews or adjusted scoring metrics – designed to promote a more equitable distribution of capital and unlock opportunities that might otherwise be overlooked, ultimately fostering a more inclusive and efficient innovation ecosystem.

The SimVC-CAS framework extends beyond simple prediction to function as a dynamic simulation environment for venture capital investment. This capability allows stakeholders – including fund managers, policymakers, and even entrepreneurs – to model hypothetical scenarios and assess the potential consequences of different interventions. For example, the framework can explore how changes to investment criteria, such as prioritizing specific sectors or demographic groups, might affect overall portfolio performance and the distribution of funding. Furthermore, it facilitates the evaluation of policies designed to mitigate bias, such as blind review processes or adjustments to scoring algorithms, providing quantifiable insights into their effectiveness before implementation. By offering this ‘what-if’ functionality, the framework empowers informed decision-making and supports the design of more equitable and efficient venture capital ecosystems.

Rigorous methodology was central to this research, with particular attention given to preventing data leakage – a common pitfall in predictive modeling. The team implemented a strict temporal split of the data, ensuring that future information never influenced predictions about the past, and employed careful feature engineering to avoid inadvertently including variables that would only be known after a startup’s outcome. This commitment to data integrity extends to cross-validation techniques, which were designed to simulate real-world investment scenarios without introducing bias. By proactively addressing these potential sources of error, the study ensures that the observed performance gains are not merely artifacts of the modeling process, but reflect a genuine ability to identify promising ventures and, crucially, that the findings can be reliably applied to unseen data and future investment decisions.

Future Horizons: Expanding the Framework and Addressing Complexity

Future research endeavors will prioritize the incorporation of temporal dynamics into the investment modeling framework. This will be achieved through the application of methods such as Graph Stream Transformer (GST), enabling the capture of evolving relationships and trends over time. Unlike static analyses, GST allows the model to consider the sequence of investments, investor behavior shifts, and the changing characteristics of startups as crucial factors influencing future funding decisions. By analyzing investment patterns as a continuous stream of events, the framework can potentially identify emerging opportunities, predict investment trajectories, and better understand how market conditions impact the flow of capital, ultimately providing a more nuanced and accurate representation of the venture capital landscape.

The current investment framework will be broadened to encompass a more nuanced understanding of diverse investment types, moving beyond a singular approach. Future iterations will specifically model the unique characteristics and risk profiles associated with seed funding, Series A rounds, and later-stage venture capital, alongside alternative investment classes. Crucially, the model will also incorporate external variables – such as prevailing market conditions, interest rates, and shifts in regulatory landscapes – to assess their influence on investment decisions and portfolio performance. This expanded scope will allow for a more realistic and predictive analysis of the venture capital ecosystem, capturing the dynamic interplay between internal investment strategies and broader economic forces.

The venture capital landscape is characterized by diverse connections between startups and investors, relationships that aren’t uniform and require nuanced modeling. To better capture this complexity, the integration of the Spatio-Temporal Heterogeneous Graph Multi-Node Neural Network (SHGMNN) promises a significant advancement. This network allows for the representation of varying relationship strengths and types – considering not just that a connection exists, but how it manifests, such as mentorship, financial backing, or strategic partnership. By accommodating these heterogeneous interactions, SHGMNN enables a more accurate prediction of investment flows and success rates, moving beyond simplistic, one-size-fits-all approaches. The network’s ability to learn from the evolving structure of these connections over time will refine the understanding of investor preferences and startup potential, ultimately leading to more informed decisions within the venture capital ecosystem.

The SimVC-CAS model, when fully implemented, demonstrated a performance level of 36.47% as measured by the F1 score. This metric represents a balanced evaluation of the model’s precision and recall in identifying relevant connections within the venture capital landscape. While acknowledging that further refinement is necessary, this initial F1 score provides a crucial baseline for assessing the model’s effectiveness and guides ongoing development efforts. The result signifies the model’s potential to discern meaningful relationships between startups and investors, offering a quantifiable step toward a more nuanced understanding of investment patterns and a foundation for future enhancements to predictive accuracy.

The culmination of this research aims to deliver a holistic platform designed to illuminate the intricacies of venture capital and facilitate its optimization. This system isn’t merely intended as a descriptive tool; it seeks to actively enhance the ecosystem by providing insights that streamline investment processes, identify promising startups with greater accuracy, and ultimately, accelerate the pace of innovation. By modeling the complex interplay between startups and investors, and accounting for dynamic market forces, the platform strives to unlock capital for ventures poised to drive significant economic growth. The anticipated impact extends beyond financial returns, fostering a more efficient and robust system capable of supporting groundbreaking advancements across diverse industries and contributing to sustained prosperity.

Our proposed method utilizes a framework integrating <span class="katex-eq" data-katex-display="false">\mathcal{F}[x]</span> to transform input <span class="katex-eq" data-katex-display="false">x</span> into a desired output <span class="katex-eq" data-katex-display="false">y</span>. — Our proposed method utilizes a framework integrating $\mathcal{F}[x]$ to transform input $x$ into a desired output $y$ .

The pursuit within SimVC-CAS, mirroring the dynamic of venture capital, inherently demands a willingness to challenge established norms. This resonates deeply with the spirit of Paul Erdős, who famously stated, “A mathematician knows a lot of things, but a good mathematician knows where to find them.” SimVC-CAS doesn’t simply accept existing investment data as truth; it actively simulates interactions, effectively ‘reverse-engineering’ the decision-making process to uncover hidden patterns and improve predictive accuracy. By modeling the co-investment network, the framework embraces the complexity of real-world investment scenarios, recognizing that truth isn’t a singular point, but an emergent property of collective behavior. The system, therefore, functions not as a passive observer, but as an active explorer of possibilities, continually testing the boundaries of prediction and interpretation.

Beyond the Simulated Deal Flow

SimVC-CAS offers a compelling glimpse into the potential of agent-based modeling for venture capital, but the true test lies not in replicating existing patterns, but in anticipating deviations. The framework currently operates within the constraints of available data – a historical echo chamber. Future iterations must deliberately introduce controlled ‘noise’ – synthetic anomalies and black swan events – to assess the robustness of the co-investment network and expose hidden vulnerabilities. The system’s predictive power is, after all, only as good as its ability to model the irrationality it seeks to understand.

A critical limitation remains the implicit assumption of investor homogeneity. While the model captures co-investment behavior, it largely treats agents as interchangeable nodes. The next phase should explore the impact of incorporating distinct ‘cognitive biases’ – genuine heuristics and flaws in judgment – into individual agent profiles. This isn’t about achieving perfect realism, but about recognizing that investment decisions aren’t purely logical; they are, fundamentally, exercises in controlled hallucination.

Ultimately, the architecture hints at a broader question: can collective simulation unlock emergent properties beyond prediction? Perhaps the most intriguing avenue lies in using the framework not just to forecast success, but to actively design it – to engineer startup ecosystems that maximize innovation and resilience. The system, after all, is a mirror reflecting not just what is, but what could be, if the rules were bent, broken, and rebuilt.

Original article: https://arxiv.org/pdf/2512.22608.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/