EventChain

Investor behavior and startup event data.

Quick start · Data model · Tables · Uses · Limitations · Citation

The dataset links investor participation to semiconductor-startup funding, operating milestones, governance changes, strategic actions, and company outcomes—without requiring a paid company database.

1,915 investors	2,547 chip startups	10,432 event rows	1,342 event chains
9 primary investor types	390 success · 942 failure · 1,215 ambiguous	funding · exposure · interface	1,071 dated windows

Stable IDs, provenance tables, schema contracts, and reusable pipeline skills make the reconstruction inspectable instead of leaving it as a collection of search results.

Why this dataset

• Behavior, not just transactions. Funding participation is linked with product, customer, supply-chain, governance, ownership, exit, and strategic events.
• Time-ordered chains. Events are assigned to funding-level outcome windows, enabling retrospective sequence and pattern analysis.
• Free-public-data reconstruction. The release is built from sources such as SEC EDGAR, Wikidata, company and investor websites, public news, and public institutional disclosures—not a paid Crunchbase export.
• Auditable by design. Entity provenance, deterministic identifiers, human-readable data dictionaries, and machine-readable Arrow contracts are included.

Quick start

Install the Hugging Face datasets library with Parquet support:

pip install datasets pyarrow

Load one table by selecting its config:

from datasets import load_dataset

chains = load_dataset(
    "quge007/eventchain",
    "chains",
    split="train",
)

funding = load_dataset(
    "quge007/eventchain",
    "funding_events",
    split="train",
)

print(chains[0])

Reproduce the release inventory, 501-company evaluable cohort, and the three primary P1--P8 action-pattern results with one local command:

python reproduce_core_results.py

The command is read-only, fails if its published reference counts drift, and prints a machine-readable JSON result. It requires only pyarrow.

There are eleven configs because the tables have different schemas. Loading them separately preserves each table's native columns and types.

Data model

investors ───────────────┐
                        ├── funding_events ──┐
associated_entities ────┤                    │
                        ├── interface_events ├── chains
chip_companies ─────────┼── exposure_events ┘
                        └── company outcome labels

companies ── categories          (broader 43,474-company entity layer)

The four chip-analysis tables use chip_companies.company_id as their company foreign key. Funding and interface investor IDs resolve to investors; resolved non-investor counterparties resolve to associated_entities. Event IDs stored inside chains resolve back to their respective event tables.

Tables

Config	Table	Rows	Row grain
investors	investors_entity_v0.3	1,915	Canonical investor
investors_provenance	investors_provenance_v0.2	4,205	Field-level source claim
chip_companies	companies_chip_subset_entity_v0.3	2,547	Canonical chip startup
funding_events	funding_events_chip_v0.3_clean_human_v0.2	3,897	Participant in a funding-level outcome
exposure_events	exposure_events_chip_v0.3_clean_human_v0.2	3,311	Externally observable company event
interface_events	interface_events_chip_v0.4_clean_human_v0.2_identity_patch_v1	3,224	Investor/company interface event
chains	chains_chip_v0.3_clean_human_v0.2_identity_patch_v1	1,342	Funding-level outcome chain
companies	companies_entity_v0.2	43,474	Canonical company in the broader entity layer
categories	category_v0.2	43,474	One category record per broader company
associated_entities	associated_entity_v0.1	1,113	Resolved non-investor counterparty
associated_provenance	associated_provenance_v0.1	3,563	Associated-entity field claim

The 3,897 funding participant rows represent 1,342 logical funding-level outcomes. The chain table contains 1,342 materialized outcome windows across 663 companies; 1,071 chains have dated windows.

Coverage and construction

The chip universe starts from 2,278 companies founded from 2010 onward and matched through OpenSporks/Crunchbase categories. Description-based recall and public industry reports expand the candidate set; scope cleaning and canonical deduplication produce the released 2,547-company universe:

• 390 SUCCESS
• 942 FAILURE
• 1,215 AMBIGUOUS

Event collection covers the 1,332 companies with explicit SUCCESS or FAILURE labels. These labels are financing/exit milestone proxies, not measurements of profitability, product quality, social value, or long-run survival.

The investor entity table contains a frozen 989-entity base plus 926 entities recovered during interface-event identity resolution. It covers institutional VC, CVC, private equity, asset managers, hedge funds, sovereign wealth funds, pension funds, investment banks, angels, and a small number of related types.

Schema and provenance

Every Parquet table has a human-readable specification under schemas/. Funding, exposure, interface, and chain tables additionally include machine-readable pyarrow-schema-json-v1 contracts defining field order, Arrow types, nullability, metadata, and controlled event enums.

Release checks enforce:

• unique entity, event, and chain primary keys;
• valid company, investor, associated-entity, and chain-event foreign keys;
• schema-version and Arrow-contract agreement;
• valid confidence ranges and parseable metadata_json;
• alignment between chain event-ID lists and their stored counts;
• label agreement between chains and the chip-company entity table.

Intended uses

This dataset is suitable for:

• retrospective investor-behavior research;
• event-sequence and event-graph methods;
• entity-resolution and provenance experiments;
• coverage and missingness analysis of public investment records;
• reproducing the released table layer and primary highlighted pattern counts of the companion research;
• teaching relational analysis with versioned Parquet data.

Out-of-scope uses

Do not use the dataset as:

• a causal estimate of whether an investor action creates success;
• a forecast of what a specific investor will do;
• a complete record of private interactions or investment amounts;
• an investment recommendation or company ranking;
• evidence that an unobserved action did not occur.

The exhaustive 23,308-pattern registry and its frozen configuration are not included in this release. The 501-company evaluable cohort and primary P2, P3, and P8 counts are deterministically reconstructed by reproduce_core_results.py; reproducing the exhaustive exploratory sweep requires the companion pattern-engine skill and frozen research registry.

Known limitations

• Public announcements overrepresent visible financing, exits, and governance changes and underrepresent private operating support.
• Freely accessible retrieval captures about 70% of the interface evidence visible in the paid-source reference; sovereign-wealth coverage is about 50%.
• Outcome-dependent date coverage means the dated-chain sample is selected; SUCCESS/FAILURE comparisons do not represent the entire chip universe.
• Historical field provenance is retained for 97 of the 989 frozen base investors; provenance for the 926 newly resolved investors is complete for their non-null adjudicated fields.
• The dataset is descriptive and observational. It does not remove residual confounding or support causal conclusions.

Sources and redistribution boundary

The reconstruction uses SEC EDGAR, Wikidata (CC0), company and investor sites, public news retrieval, AKShare, S-1 prospectuses, 13D/13G filings, 13F holdings, and public institutional disclosures.

Where Crunchbase categories are used, they come from the OpenSporks free database snapshot downloaded from Hugging Face and labeled MIT, supplemented by then-public information. Because the underlying records originate from Crunchbase, this release does not redistribute the raw OpenSporks/Crunchbase mirror rows. It contains structured derived fields, matched category tags with source labels, and self-produced analytical tables.

The released dataset is licensed under CC BY 4.0. Users remain responsible for respecting any restrictions that apply when following evidence links back to their original third-party sources.

Companion resources

• Reusable pipelines and agent skills: quge009/startup-decision-skills
• Companion paper: From Ideas to Actions: A Public-Data Decision-Support Toolchain Across the Venture Lifecycle. SSRN | arXiv.

Citation

Please cite the dataset and its companion paper: SSRN | arXiv. The paper entries below identify the same work on two platforms; either may be used. Machine-readable citation metadata is available in CITATION.cff.

@dataset{qu2026investorbehavior,
  author    = {Lei Qu},
  title     = {EventChain: Investor Behavior and Startup Event Data},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/quge007/eventchain}
}

@techreport{qu2026ideastoactions,
  author      = {Lei Qu},
  title       = {From Ideas to Actions: A Public-Data Decision-Support Toolchain Across the Venture Lifecycle},
  year        = {2026},
  institution = {SSRN},
  number      = {7445800},
  doi         = {10.2139/ssrn.7445800},
  url         = {https://ssrn.com/abstract=7445800}
}

@misc{qu2026ideastoactions_arxiv,
  author      = {Lei Qu},
  title       = {From Ideas to Actions: A Public-Data Decision-Support Toolchain Across the Venture Lifecycle},
  year        = {2026},
  eprint      = {2609.15219},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  doi         = {10.48550/arXiv.2609.15219},
  url         = {https://arxiv.org/abs/2609.15219}
}