Insider Trading Controls For Alternative Data Usage

When to Use

Use this skill at the point where an already-onboarded alternative dataset (satellite imagery, credit card transactions, web-scraped text, consumer geolocation) is about to drive trading, and on every re-check thereafter. It converts a dataset's compliance posture into an auditable AltDataComplianceReport recording, control by control, what was checked and what failed.

The controls support an investment adviser's Section 204A obligation to maintain and enforce written policies reasonably designed to prevent misuse of material non-public information. That obligation is independently chargeable: the SEC has settled Section 204A actions where the orders contained no finding that anyone actually traded on MNPI — the deficient policy was the violation.

When NOT to Use

• Not for vendor onboarding triage. Legal-rights, CFAA/scraping posture, ToS review, and anonymisation methodology belong to alternative-data-vendor-due-diligence-checklist, which emits the DiligenceRecord this engine expects to already exist. This skill consumes that outcome as the has_vendor_diligence_signoff input; it does not re-derive it.
• Not a legal determination. Insider-trading liability turns on breach of a duty of trust or confidence (United States v. O'Hagan, 521 U.S. 642 (1997); 17 CFR 240.10b5-2), not on a boolean. A LOW_RISK_APPROVED verdict is evidence for counsel, not a substitute for counsel.
• Outside the US. EU MAR (Art. 7, 8, 14) defines inside information and the prohibition differently and has no Section 204A analogue; see eu-market-abuse-regulation-mar-surveillance. UK, Singapore and Hong Kong regimes are not modelled.
• Not an issuer-level restricted-list check. This gate scores a dataset. Whether a specific issuer is restricted, and whether the firm holds MNPI from a non-alt-data channel, is a separate control.
• Not a data-quality or lookahead gate. See backtesting-alt-data-strategies-with-realistic-availability-lag.

Prerequisites

• A completed AltDataDatasetSpec. Every boolean must be a real bool backed by evidence the firm holds, not a vendor assertion — the engine raises AltDataComplianceError on a coerced value rather than scoring it.
• A current vendor diligence record from alternative-data-vendor-due-diligence-checklist.
• A calibrated, written threshold policy. The defaults (min_panel_aggregation_count=50, earnings_blackout_window_hours=48.0) are engineering defaults with no regulatory basis — no regulator prescribes either number. See references/standards.md before adopting them.

Workflow

1. Set has_mnpi_risk on provenance, not on predictive power. Alt data is nonpublic and material by design — that is why it is bought. The SEC staff's own position is that alternative data "does not necessarily contain MNPI." The flag belongs to a different question: does the dataset's origin imply a breached duty of trust or confidence — leaked or hacked material, data supplied to the vendor under confidentiality obligations, an insider or tippee source? Where provenance is genuinely unknown, default to True and escalate; do not guess in the permissive direction.
2. Carry the vendor diligence outcome, and separate it from ToS. has_vendor_diligence_signoff records that a current diligence record exists; is_tos_compliant records that collection is consistent with the source's terms. They fail into distinct failed_controls entries because the exposures differ: a missing sign-off is a Section 204A policy failure, while a terms breach is contract/tort exposure and can itself supply the confidentiality duty whose breach makes the data MNPI. Re-run the gate whenever the vendor changes collection methodology — the 2022 Risk Alert cites failure to define when re-diligence is required as a named deficiency.
3. Score PII scrubbing and panel aggregation as two controls. is_pii_scrubbed asks whether identifiers were removed and that removal verified; panel_aggregation_count is a k-anonymity-style cell-size floor. A 500,000-contributor panel that was never scrubbed fails scrubbing — the audit note must say so rather than blame the panel size.
4. Apply the earnings blackout as declared firm policy. hours_to_earnings_release is a signed distance: positive before the release, negative after. The gate compares its absolute value, so the window is two-sided. Pass None when no release is scheduled in the monitoring horizon.
5. Classify, then read the detail, not just the verdict. All four controls are evaluated unconditionally before classification, so every is_*_cleared flag is a tested result. risk_classification reports the most severe failure; failed_controls reports all of them. Persist the whole report — the record is the point.

Full procedure: see references/workflows.md. Standards reference: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

• Treating a vendor's written representation as the control. App Annie's Terms of Service promised that developer metrics would be used only in aggregated, anonymised form; the SEC found it used non-aggregated, non-anonymised data anyway to sharpen estimates for trading clients (Admin. Proc. 34-92975, 2021-09-14; $10M and $300k penalties, the first securities-fraud action against an alt-data provider). Obtaining the representation is necessary and insufficient. Require independent verification — right-to-audit exercised, sample-data inspection, or third-party attestation.
• Mapping compliance answers through a string. A spec assembled from CSV, JSON, or an LLM tool call carries 'no' and 'false' strings. Every non-empty string is truthy, so a naive engine reads is_pii_scrubbed='no' as a pass on the very control it exists to enforce. This engine raises rather than coerces. Do not add a bool() cast at the call site to make the error go away.
• Reporting a clearance you never tested. A rejection on one gate must not hard-code the others as cleared. An MNPI-rejected dataset one hour from an earnings release is not outside the blackout window, and a record saying otherwise is a false statement in a compliance file.
• Presenting the thresholds as legal requirements. Neither GDPR Recital 26 nor CCPA/CPRA § 1798.140 specifies a minimum group size, and no rule imposes an alt-data earnings blackout at all — the only codified waiting periods in this area are the Rule 10b5-1(c) cooling-off periods (90 days for directors and officers, 30 for others), which govern insider trading plans, not research data. Documenting N ≥ 50 as "required by GDPR" misstates the law to your own auditors and to any regulator reading the file.
• Assuming lawful scraping means unrestricted scraping. After Van Buren v. United States, 593 U.S. 374 (2021) and the hiQ v. LinkedIn line, scraping public pages without authentication is likely not a CFAA violation — but hiQ still lost on breach of contract, settling in December 2022 with a $500,000 judgment and admitted liability for trespass to chattels. "Not a federal crime" is not "no exposure."
• Reading BLACKOUT_WINDOW_RESTRICTED as a rejection. It is a time-boxed pause on an otherwise-clean dataset; it clears on its own. Wiring it to the same remediation path as REJECTED_MNPI_RISK buries genuine rejections in routine noise.

Verification

• Audit a compliant satellite dataset (no MNPI risk, diligence signed off, ToS clean, PII scrubbed, panel 250, 72h to earnings) and confirm LOW_RISK_APPROVED with failed_controls == ().
• Set has_mnpi_risk=True on a spec that is also 1h from earnings with panel_aggregation_count=1, and confirm REJECTED_MNPI_RISK with is_blackout_window_cleared and is_pii_anonymization_cleared both False — no untested gate may report a pass.
• Set is_pii_scrubbed=False with panel_aggregation_count=500_000 and confirm the note names PII scrubbing and does not claim the panel is below the minimum.
• Set only is_tos_compliant=False and confirm failed_controls == ('TERMS_OF_SERVICE',), distinguishable from a missing sign-off.
• Pass has_vendor_diligence_signoff='no' and confirm AltDataComplianceError — not LOW_RISK_APPROVED.
• Check the boundaries: panel_aggregation_count=50 clears and 49 fails; hours_to_earnings_release=48.0 clears and 47.9 restricts; -6.0 restricts exactly as +6.0 does.
• Pass float('nan') or a negative panel count and confirm AltDataComplianceError rather than a scored report.
• Run python -m unittest discover -s skills/insider-trading-controls-for-alternative-data-usage/scripts and confirm a 100% pass rate.

Related Skills

• alternative-data-vendor-due-diligence-checklist
• alternative-data-feature-integration
• backtesting-alt-data-strategies-with-realistic-availability-lag
• web-scraped-sentiment-data-pipeline
• eu-market-abuse-regulation-mar-surveillance