The IP-risk dataset — point-in-time, survivorship-bias-free
Every red flag is dated as of the day it became a public USPTO record, evidence-backed and Ed25519-signed. Entities that later went bankrupt or delisted stay in the set — no survivorship bias, the failures are the ground truth. The accumulating point-in-time history is the asset; the live feed is the hook.
What's in it
Patent security-interest events (companies pledging patents as loan collateral), 1980–2024, growing daily — plus portfolio transfers, maintenance lapses, PTAB challenges and abandonment clusters as coverage expands.
Per event: entity, flag code, severity 0–100, confidence, as_of, evidence hash + USPTO source link, rule version, and a signature.
A dated SEC ticker/CIK crosswalk (valid_from/valid_to) so you can join to public-filer identity without re-introducing look-ahead.
Why it's defensible
Point-in-time strict — no look-ahead; every feature is computed as-of a timestamp.
Survivorship-bias-free — the bias hits distressed small-caps hardest, exactly this universe; a quant rejects a set that silently drops the failures.
Reproducible — each event replays its evidence path (filing → field → hash → parser + rule version). See the backtest for the signal→distress lead-time study.
What a sample looks like
Each row is one red flag, in JSONL and CSV, shipped with a signed MANIFEST.json (per-file SHA-256, the flag/rule-version inventory, the as_of span, and the Ed25519 signing identity). Per row: entity (+ point-in-time SEC ticker/CIK), flag, severity 0–100, confidence, as_of, evidence_hash, source URL, rule version, and a signature you can verify against /v1/pubkey. Every field is defined in the data dictionary.
Access
Research sample + data dictionary — one-off, to evaluate coverage and schema before licensing.
Full historical PIT license — the complete signed history with updates, annual.
API + webhooks — real-time programmatic access; see pricing.