An AI that listens to expert calls improves startup selection: a model combining LLM-extracted signals with sequential Bayesian updating raises prediction accuracy by 6.7% and boosts backtested portfolio ROI by 15.3%, disproportionately helping young, tech‑intensive, diverse, and low‑visibility startups.

Main Finding

A sequential LLM–Bayesian framework that converts expert-network call transcripts into time-varying beliefs and uncertainty materially improves measurement of investor learning and materially improves investment outcomes in private markets. Expert calls contain decision-relevant soft information: a single call raises subsequent deal probability by ~6.9–9.0 percentage points; positive sentiment raises deal likelihood by ~3.9–4.1 pp; topic-specific discussions (technology adoption, customer acquisition) can raise deal probability by up to ~14.7 pp. Modeling calls as sequential signals (LLM extraction + Bayesian belief updates) yields economically meaningful gains: portfolio ROI increases by ~15.26% and predictive F1 by ~6.69%, with most gains concentrated in the upper tail of returns.

Key Points

• Data scope: large dataset of expert-network call transcripts matched to startups (2017–2024). Average transcript ≈ 5,880 words and 74 Q–A exchanges. Firms typically: >50 employees, ~5 financing rounds, ~14 institutional investors.
• Framework: an LLM extracts structured signals (content, sentiment, topic, speaker type) from unstructured multi-turn dialogues; a Bayesian network with latent variables performs sequential belief updating (time-varying success belief + uncertainty).
• Economic effects (reduced-form):
  • Any expert call → +6.9–9.0 pp deal probability.
  • Positive sentiment → +3.9–4.1 pp deal probability.
  • Technology adoption and customer-acquisition discussions → up to +14.7 pp (especially in high-uncertainty settings).
• Belief mechanism:
  • Raw text signals lose explanatory power once model-implied beliefs and uncertainty are included.
  • A 1 standard-deviation increase in inferred success belief → ≈ +11 pp deal probability.
  • Reductions in uncertainty (learning) also raise investment likelihood.
• Temporal and horizon asymmetries:
  • Positive signals better predict short-term investment behavior.
  • Negative signals are more informative of long-run firm performance/failure.
• Heterogeneity and state dependence:
  • Signals are most valuable for technologically complex firms, early-stage/young firms, low-visibility firms, and diverse founding teams—i.e., where information frictions are severe.
  • Attention weights decline with firm maturity; operational sources (customers, consultants, former execs) receive highest weights.
  • In tech domains, positive signals are weighted 2–6× more than negatives; attention to customer acquisition rises markedly for positive cues (example attention: 0.023 negative → 0.102 positive).
• Model performance and robustness:
  • Modeling sequential arrival and dependencies matters: ablation (removing within-call or across-call dependencies or belief updating) substantially reduces predictive and economic performance.
  • Adding more calls improves accuracy (from 0.769 → 0.777) and ROI (over +32.5% reported in progressive incorporation tests).
  • Gains persist across a variety of payoff and cost assumptions and vs. strong baselines (standard ML, transformer-based models, and recent text approaches).
• Economic allocation effect: reallocating capital toward firms ranked highly by the model systematically increases ROI and MOIC, concentrating improvements in the upper tail of returns.

Data & Methods

• Data:
  • Expert network call transcripts, 2017–2024, compliance-archived.
  • Matched firm fundamentals from Crunchbase and VentureXpert.
  • Sample restricted for analysis (e.g., firms with >50 employees) to focus on commercially meaningful startups.
• Signal extraction:
  • LLMs (used as structured-text parsers/classifiers) extract per-call features: topic labels, sentiment/polarity, expert type, and finer conversational cues (QA structure).
• Sequential belief model:
  • Bayesian network with latent state variables representing investor belief about firm success and associated uncertainty.
  • Treats each call as a sequential signal whose credibility depends on source/topic/context; updates beliefs over time using Bayesian updating rules.
• Evaluation:
  • Predictive metrics: accuracy, F1-score (F1 improved up to +6.69% vs. best baselines).
  • Economic metrics: portfolio ROI, MOIC, tail-return analyses (ROI improvement ~+15.26% when using the model for allocation).
  • Ablation studies: remove dependencies, remove belief updating, or remove sequential structure to isolate contribution of each component.
  • Event-study / reduced-form regressions to show call activity spikes before deals and to quantify the marginal effect of calls and sentiment on deal probability.
  • Attention and interpretability analyses to identify which sources/topics drive belief updates and how these vary with firm state.
• Identification caveats:
  • Calls are investor-initiated (endogenous). The paper addresses endogeneity via event-study patterns, state-dependent analyses, and by showing that model-implied beliefs mediate raw-signal effects—but observational limitations remain (no randomized assignment of calls).

Implications for AI Economics

• Methodological:
  • Integrating LLM-based signal extraction with explicit sequential Bayesian updating creates a practical pipeline for converting noisy, high-dimensional qualitative text into economically interpretable belief trajectories and uncertainty measures. This hybrid approach outperforms black-box text models where sequential signal structure matters.
  • Modeling information arrival and credibility (source/topic/context) is crucial—representation improvements alone (e.g., transformers) are insufficient when decisions depend on how information accumulates.
• Financial economics and markets:
  • AI tools can materially reduce information frictions in opaque markets (early-stage venture investing) by quantifying soft signals and their value, improving capital allocation and increasing realized returns—especially in high-uncertainty domains.
  • The asymmetric weighting of positive and negative signals and the horizon-dependent informativeness (positives → short-term, negatives → long-term) suggest potential behavioral and market-structure consequences (e.g., overemphasis on short-term validation, underweighting downside), which AI systems may amplify unless explicitly corrected.
• Policy, governance, and risk:
  • Widespread adoption of LLM-Bayesian tools raises governance questions: model auditability, transparency of attention/weights, and regulatory compliance (expert calls are compliance-sensitive). Policymakers and platforms should consider standards for explainability and data handling to prevent misuse or inadvertent leakage of material non-public information.
  • Reliance on LLM-extracted signals introduces model risk: LLM biases, prompt-engineering choices, and training-data artifacts can affect inferred beliefs. Continuous validation (including human-in-the-loop checks) and robustness testing across LLM versions are important.
• Market dynamics and second-order effects:
  • If many investors adopt similar LLM-based learning systems, information aggregation may change—potentially increasing market concentration (more capital to model-identified winners) and altering incentives for startups and experts (gaming signals, staged disclosures).
  • The highest gains accrue where traditional data are weakest (complex tech, early stage), implying AI could reshape which projects get funded and how due diligence resources are allocated.
• Research directions:
  • Causal tests: randomized experiments (e.g., RCTs assigning calls or summaries) to more cleanly identify causal effects of information acquisition.
  • Generalizability: apply the sequential LLM–Bayesian framework to other domains with soft information (M&A due diligence, credit underwriting, policy deliberations).
  • Robustness to LLM choice and prompt design: evaluate how different LLMs or extraction schemes change belief estimates and economic outcomes.
  • Dynamic feedback: study how the presence of automated belief-measurement tools affects expert behavior, startup disclosure strategies, and market equilibria.

Summary takeaway: The paper provides strong empirical evidence that LLMs combined with sequential Bayesian belief-updating can turn conversational soft information into actionable, measurable beliefs and uncertainty estimates, improving investment decisions in opaque private markets—while highlighting important behavioral, governance, and research implications for AI-driven information processing in finance.