0 cumulative citations
View corpus contextA reproducible, public-data decision toolchain can moderately predict startup viability and reconstruct investor–company event chains: deterministic pre-founding checks achieve a combined F0.5 ≈ 0.63 on held-out samples, and post-investment public records show sustained operating progress (not financing alone) aligns with better observed outcomes, though coverage gaps and retrospective labels limit causal claims.
From Ideas to Actions: A Public-Data Decision-Support Toolchain Across the Venture Lifecycle
The paper presents a reproducible public-data toolchain that produces deterministic pre-founding viability verdicts with modest predictive performance (combined F0.5 ≈ 0.63) and an auditable EventChain reconstruction showing that sustained product/customer/supply-chain progress (not just financing) correlates with better post-investment outcomes, while visible post-investment activity is dominated by follow-on financing.
Citation observations
Cumulative provider counts captured on specific dates; providers are never combined.
Founders face two linked decisions: whether to pursue an idea before founding, and which operating actions and capital partners fit afterward. We present a public-data decision-support toolchain combining time-bounded proposal profiling, market and moat checks, and deterministic aggregation with auditable investor-company event chains for retrospective analysis. Pre-founding: (a) After threshold selection on 198 development companies, the frozen pipeline achieves F0.5=0.5357 [0.412, 0.655] on an independent, row-disjoint 198-company validation sample. On the combined 396 rows, the Full Pipeline scores 0.6301 versus 0.2734 for a paired Raw LLM baseline. Post-stratification of 1,027 completed cases in a separate scale cohort yields 0.6506 [0.598, 0.707]; the run remains incomplete. A 377-row composition-matched check yields 0.6573. (b) The AI-inference study identifies distribution-layer businesses as a replicable path to independent profitability with a limited revenue ceiling, and frontier-model ownership as a path to capital-market upside at exceptional capital cost. Post-founding: (a) Public sources support auditable event-chain analysis. (b) In the chip-company study, sustained product, customer, and supply-chain progress is associated with better observed outcomes; financing alone does not establish operating progress. (c) Financing comprises 79% of confirmed visible post-investment actions. Evidence tentatively favors acquisition-experienced strategic corporate investors for acquisition-oriented founders and financing-led institutional VCs with fewer observed control events for independence-oriented founders. Findings are developmental and observational, not causal guarantees or investment advice. We release shared ontology, provenance-bearing EventChain data, schemas, benchmarks, and executable skills for audit, reuse, and extension.
Summary
Main Finding
The paper builds and validates an end-to-end, reproducible decision-support toolchain—entirely from free public sources—that (1) evaluates idea-stage proposals via time-bounded market and moat checks plus a deterministic aggregator, and (2) reconstructs provenance-bearing investor–company event chains to analyze post-investment operating patterns. The pre-founding pipeline achieves materially positive predictive performance versus raw-LM baselines across independent holdouts (F0.5 ≈ 0.53–0.66 depending on sample and post-stratification), and the post-founding EventChain analyses show that observable continued financing dominates recorded post-investment actions (≈79%) while operational progress (product, customer, supply-chain signals) correlates with better outcomes — financing participation alone does not.
Key Points
- Two-part public-data toolchain:
- Part One (pre-founding): converts a proposal into a five-dimension candidate card, runs an M-check (time-bounded TAM/SAM/SOM + Porter five forces) and a U-check (moat via Helmer’s Seven Powers + VRIO), then deterministically aggregates into PASS/WARN/FAIL (score = 0.5M + 0.5U; thresholds fixed).
- Part Two (post-founding): builds versioned, provenance-bearing EventChains from SEC filings, portfolio pages, news, etc., enumerates label-blind dated patterns, and analyzes retrospectively observed investor/company behaviors.
- Candidate-card five dimensions: revenue model, customer segmentation, cost structure, differentiation & moat (Helmer’s Seven Powers), and strategic vulnerabilities.
- Data sources (public-only): SEC EDGAR (Form D, 13D/13G), Wikidata, OpenSporks (MIT-licensed Crunchbase snapshot), company portfolio pages, S-1s, 13F holdings, national-fund disclosures, news APIs. All inputs are provenance-linked and time-bounded (e.g., market evidence restricted to three years before founding through founding year).
- Performance benchmarks (retrospective evaluation against deterministic SUCCESS/FAILURE outcome ontology based on financing/exits):
- Development/validation: frozen system F0.5 = 0.5357 [0.412, 0.655] on an independently drawn, row-disjoint 198-row validation sample; combined 396-row benchmark F0.5 = 0.6301.
- Raw LLM baseline on same 396 rows: F0.5 = 0.2734. (Other external baselines such as GPT-4o / VCBench not directly comparable due to different tasks/datasets.)
- Scale check (incomplete run): post-stratified 1,027 completed cases → F0.5 = 0.6506 [0.598, 0.707]; a 377-row composition-matched check → F0.5 = 0.6573.
- Post-founding findings:
- Public, announcement-level records cover ≈70% of investor–company interface events; monetary amounts are populated for ≈16% of rows; sovereign-wealth coverage ≈50%.
- Confirmed publicly visible post-investment actions are dominated by continued financing (≈79%).
- In a chip-company implementation, sustained product, customer, and supply-chain progress correlate with better outcomes; financing participation alone is a weak proxy for ongoing operational progress.
- Tentative recommendation patterns: founders aiming for acquisition may prefer strategic corporate investors with relevant acquisition histories; independence-oriented founders may prefer institutional VCs with fewer observable control events. These are observational associations, not causal prescriptions.
- Reproducibility and limits:
- The authors release a reproducibility stack (shared ontology, EventChain data, schemas, benchmarks, executable skills) but do not redistribute vendor-sourced row-level data that would violate terms.
- The work is explicitly observational and methodological: it documents associations and builds auditable artifacts, not causal claims nor investment advice.
Data & Methods
- Outcome ontology: deterministic company labels built from structured records (SUCCESS vs FAILURE proxies). Fine-grained labels (IPO, late-stage funded, acquired, delisted, closed, no-traction, indeterminate/unknown) map to three-level outcomes used for scoring; AMBIGUOUS/UNKNOWN excluded from binary evaluations.
- Pre-founding pipeline details:
- Candidate profiler: creates a structured five-section card from proposal text, category tags, tagline, and website text; leakage-scan removes retrospective outcome cues.
- M-check: retrieves time-bounded market evidence, computes TAM/SAM/SOM headroom and discounts with Porter five forces and incumbent-lock factor to produce a verdict plus evidence JSON.
- U-check: assesses claimed moats using Helmer’s Seven Powers and a VRIO lens producing a verdict plus evidence JSON.
- Aggregator: maps M/U verdicts to numeric scores (PASS=1.0, WARN=0.5, FAIL=0.0) and computes 0.5M + 0.5U with fixed cutoffs for PASS/WARN/FAIL. Only FAIL+FAIL produces final FAIL.
- Benchmarking protocol:
- Uses a development 198-row tuning cohort, an independently drawn 198-row validation cohort disjoint by row, and combined 396-row benchmark for primary reporting.
- Applies leakage controls, bootstrap/Wilson CIs for F0.5, and separate larger-scale stratified checks (1,027 completed cases with post-stratification and a 377-row matched subset).
- EventChain construction:
- Fuses dated events from EDGAR, portfolio pages, S-1s, news, and other public artifacts into provenance-tagged investor→company event graphs.
- Pattern enumeration: label-blind grammar generates chain patterns (23,308 patterns enumerated) and catalogs retrospective associations between patterns and company outcomes.
- Confirmed-post-investment analyses restrict to interface events strictly after investment to avoid leakage.
- Coverage limitations explicitly measured (e.g., ~70% event coverage) and documented.
Implications for AI Economics
- Feasibility of public-data decision tools: The study demonstrates that materially useful founder- and analyst-facing decision signals can be constructed from free, auditable public sources—lowering barriers to independent research and partially reducing information asymmetry in venture markets.
- Evidence on AI-inference business models: The AI-inference market case-study highlights a structural bifurcation:
- Distribution-layer plays (infrastructure, orchestration, distribution) are more replicable paths to independent profitability but face natural revenue ceilings.
- Frontier-model ownership (training/owning top models) potentially delivers larger capital-market upside but requires exceptional capital and entails higher risk/fragility. This informs capital-allocation trade-offs in AI economics: expected return profiles differ substantially by business model class and capital intensity.
- Measuring investor behavior and value-add: Public-event reconstructions can reveal heterogeneity in ex-post investor behavior (e.g., control events, exits, board involvement). That makes it possible to align investor selection to founder objectives (acquisition vs independence) using observable historical associations—though these are not causal and are limited by public-coverage gaps.
- Operational signals matter beyond financing counts: For capital-market inference and policy, the finding that sustained product/customer/supply-chain progress correlates with better outcomes (while financing alone does not) underscores the need for operational indicators (traction, supply integration) in models of startup value creation, especially for hardware- and deep-tech-intensive AI firms.
- Research & policy directions:
- Public-data toolchains facilitate reproducible, audit-friendly economic analyses of venture ecosystems and can enable more transparent benchmarking of business-model viability and investor conduct.
- To inform causal policymaking or investment rules, follow-up work should combine richer (possibly paid) data, improved coverage, and causal identification strategies (instrumental variables, natural experiments, or randomized interventions).
- Scaling these methods could help regulators and public funds assess ecosystem outcomes without complete reliance on proprietary datasets, but careful attention to selection bias and missingness is required.
Caveats: results are observational and methodologic, not causal; outcome labels are financing/exit proxies and not measures of profitability or social value; public-data coverage is incomplete and limits generalization; this is not investment advice nor an investor ranking.
Assessment
Paper Typedescriptive
Evidence Strengthmedium — The paper provides retrospective validation with held-out, row-disjoint samples and bootstrap CIs (development 198 rows; independent 198-row validation; combined 396-row benchmark; separate 1,027-row scale cohort and a 377-row composition-matched check). It documents leakage controls and provenance, and reports performance metrics (e.g., combined F0.5 ≈ 0.63). However, the scale validation is incomplete, outcome labels are financing/exit proxies rather than direct economic performance, and public-data coverage limitations (e.g., ~70% announcement coverage for interface events, limited per-investor amount data) constrain inference and external validity.
Methods Rigormedium — The methods show careful design: deterministic aggregation, explicit leakage-control steps, provenance-bearing data, row-disjoint validation, bootstrap CIs, and clear documentation of data sources and limitations. Nevertheless, the benchmark relies on retrospective proxies for pre-founding inputs, the sample selection and coverage of free public sources induce nonrandom missingness, the scale-validation run is incomplete, and the analyses are descriptive/associational rather than causal.
SamplePart One: development cohort of 198 companies (threshold tuning), independent row-disjoint validation of 198 companies (combined 396-row benchmark), separate 1,027-row completed-case scale cohort (incomplete run), and a 377-row composition-matched check; a reserved 1,199-row split remains untouched. Part One uses public sources (OpenSporks Crunchbase snapshot, company websites, SEC EDGAR filings, news APIs, Wikidata) and evaluates against a deterministic outcome ontology mapping financing/exit events to SUCCESS/FAILURE. Part Two: EventChain dataset reconstructed from public filings, portfolio pages, news and other free sources; an AI-inference vendor study of 29 vendors; coverage statistics reported (e.g., interface events ~70% covered, per-investor amounts ~16% populated, sovereign-wealth coverage ~50%).
Themesinnovation governance
GeneralizabilitySelection bias toward firms and events visible in free public sources (firms without public filings or news are undercovered)., Outcome proxy is financing/exit milestone (SUCCESS/FAILURE), which does not measure profitability, productivity, or broader economic impact., Retrospective candidate cards may not fully reproduce true idea-stage information available to a founder at founding (possible retrospective leakage despite controls)., Scale-validation incomplete and sample sizes modest for some checks, limiting confidence in deployment-grade generalization., Industry case studies (AI-inference: 29 vendors; chip companies) are narrow and may not generalize across sectors or geographies., Likely geographic and regulatory skew from reliance on EDGAR and other US-centric public filings, unless explicitly broadened.
Claims (11)
| Claim | Direction | Outcome | Confidence & Evidence | Details |
|---|---|---|---|---|
| After threshold selection on a 198-company development sample, the frozen proposal-evaluation system achieved an F0.5 score of 0.5357 on an independently drawn, row-disjoint 198-company validation sample. Decision Quality | positive | F0.5 performance in predicting the paper's company funding/exit outcome proxy |
Reading fidelity
high
Study strength
high
|
n=198
F0.5 = 0.5357 [0.412, 0.655]
|
| Across the combined 396-company benchmark, the proposal-evaluation system achieved an estimated F0.5 of 0.6301, compared with 0.2734 for a paired raw LLM baseline. Decision Quality | positive | F0.5 prediction performance |
Reading fidelity
high
Study strength
medium
|
n=396
F0.6301 versus 0.2734
|
| In a separately drawn scale cohort, post-stratifying all 1,027 completed cases to the fixed execution-split composition produced an F0.5 of 0.6506, while a 377-row composition-matched check produced an F0.5 of 0.6573. Decision Quality | positive | F0.5 prediction performance in scale-cohort evaluation |
Reading fidelity
high
Study strength
medium
|
n=1027
F0.6506 [0.598, 0.707]; composition-matched check F0.6573
|
| The authors state that the incomplete scale run prevents a deployment-grade generalization claim for the proposal-evaluation system. Decision Quality | negative | Generalizability to deployment conditions |
Reading fidelity
high
Study strength
high
|
not reported
|
| In the AI-inference market study, distribution-layer businesses are described as offering the most replicable route to independent profitability, but with a limited revenue ceiling, whereas frontier-model ownership offers greater capital-market upside at exceptional capital cost. Firm Revenue | mixed | Independent profitability and capital-market upside of business-model paths |
Reading fidelity
high
Study strength
medium
|
n=29
|
| Public filings, portfolio pages, company disclosures, and news can support reproducible analysis of post-investment company and investor activity without subscription-only venture data. Organizational Efficiency | positive | Reproducibility and feasibility of public-data venture analysis |
Reading fidelity
high
Study strength
medium
|
not reported
|
| In the chip-company implementation, sustained product, customer, and supply-chain progress is associated with better observed company outcomes. Firm Productivity | positive | Observed company outcomes associated with product, customer, and supply-chain progress |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Financing participation alone does not establish continuing operating progress in the chip-company implementation. Firm Productivity | null_result | Continuing operating progress following investor financing participation |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Continued financing accounts for 79% of confirmed publicly visible post-investment actions. Task Allocation | positive | Distribution of confirmed publicly visible post-investment actions |
Reading fidelity
high
Study strength
medium
|
79%
|
| The observational evidence tentatively favors strategic corporate investors with relevant acquisition histories for acquisition-oriented founders, and financing-led institutional VCs with fewer observable control events for independence-oriented founders. Task Allocation | mixed | Fit between investor behavior and founder objectives concerning acquisition or independence |
Reading fidelity
high
Study strength
low
|
not reported
|
| Public announcement-level records cover approximately 70% of interface events between investors and companies, provide per-investor amounts for only about 16% of rows, and cover sovereign-wealth investors at approximately 50%. Organizational Efficiency | negative | Coverage and completeness of publicly observable investor-company event data |
Reading fidelity
high
Study strength
high
|
∼70% event coverage; ∼16% of rows with per-investor amounts; ∼50% sovereign-wealth coverage
|