Key Takeaways
- • Machine learning for venture capital should organize signals around a defined investment decision rather than serve as a verdict on a startup, according to Rebellion Research's analysis.
- • The article cites Affinity data showing the share of venture capital firms using AI in investment decisions rose from 13% to 28% in 2026.
- • Rules-based screens, supervised learning and clustering each answer different questions and carry distinct data requirements and limitations.
- • Models must be validated against information available at the decision date to avoid leakage, survivorship bias and selective outcome labels.
- • Adoption should proceed in stages: define one decision, audit data, test retrospectively, pilot, and assess workflow value alongside decision quality.
Summary
Rebellion Research argues that machine learning is most useful in venture capital when each analytical signal addresses a specific question, such as which companies merit a first meeting or what evidence justifies deeper diligence. The article cites Affinity data showing that the share of venture capital firms using AI in the investment decision process rose from 13% to 28% in 2026, while cautioning that additional analysis does not automatically improve decisions. Without a clear link to a decision, fragmented startup data can yield confident-looking scores built on noise or bias.
The piece distinguishes among sourcing, screening, diligence, portfolio monitoring and exit analysis, noting these are different tasks rather than one universal prediction problem. It emphasizes that private-company data is often uneven in coverage and cadence, that definitions of metrics such as revenue or customers can vary across companies, and that missing data may reflect limited disclosure rather than weak performance. A model signal may take the form of a ranking, a probability estimate or an anomaly flag, and each depends on its target, inputs and the date those inputs were available. The article characterizes model outputs as conditional evidence that should prompt inquiry rather than close it.
The recommended workflow runs from defining the decision, to assembling data with preserved provenance and timestamps, constructing features, validating against historical information available at the time, and reviewing outputs in context. The article warns that training labels such as later funding rounds or exits reflect selection and survivorship effects, and that public web signals such as website changes or job listings are leads rather than verified performance facts.
A comparison table contrasts three approaches. Rules-based screening answers which companies meet explicit criteria but may encode arbitrary thresholds. Supervised learning estimates resemblance to historical outcomes but depends on sparse, selective or unstable labels. Clustering and similarity methods group companies for exploration but do not establish investability. The article concludes that a simple screen may outperform a complex model when outcomes are sparse, labels are inconsistent or market conditions have shifted.
The piece distinguishes among sourcing, screening, diligence, portfolio monitoring and exit analysis, noting these are different tasks rather than one universal prediction problem. It emphasizes that private-company data is often uneven in coverage and cadence, that definitions of metrics such as revenue or customers can vary across companies, and that missing data may reflect limited disclosure rather than weak performance. A model signal may take the form of a ranking, a probability estimate or an anomaly flag, and each depends on its target, inputs and the date those inputs were available. The article characterizes model outputs as conditional evidence that should prompt inquiry rather than close it.
The recommended workflow runs from defining the decision, to assembling data with preserved provenance and timestamps, constructing features, validating against historical information available at the time, and reviewing outputs in context. The article warns that training labels such as later funding rounds or exits reflect selection and survivorship effects, and that public web signals such as website changes or job listings are leads rather than verified performance facts.
A comparison table contrasts three approaches. Rules-based screening answers which companies meet explicit criteria but may encode arbitrary thresholds. Supervised learning estimates resemblance to historical outcomes but depends on sparse, selective or unstable labels. Clustering and similarity methods group companies for exploration but do not establish investability. The article concludes that a simple screen may outperform a complex model when outcomes are sparse, labels are inconsistent or market conditions have shifted.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.