Phase 2: Forensic evaluator, Beneish, STA/SNOA/COMBOACCRUAL

[JLaborda]

Parent

#85

PRD: spec/prds/phase2/prd.md · Spec: spec/features/008-permanent-loss-filter/spec.md

Spec

spec/features/008-permanent-loss-filter/spec.md

Scope (expanded)

Implement the forensic evaluator as a hard-exclusion module before QV scoring:

1. Distress / bankruptcy hard rules (BK_*): Altman Z, interest coverage, net debt/EBITDA, negative equity, delisted
2. Beneish M-Score (PROBM in Quantitative Value Ch. 3) with bottom-5% cross-sectional gate (FRD_BENEISH_BOTTOM_PCT)
3. STA (scaled total accruals) and SNOA (scaled net operating assets) with COMBOACCRUAL bottom-5% gate (FRD_COMBOACCRUAL_BOTTOM_PCT) — pulled forward from phase 2b per product decision (July 2026)
4. EDGAR confirmed-fraud rules coded as unavailable until SEC ETL is wired
5. Output curated/permanent_loss/run_date=.../exclusions.parquet + review queue
6. CI regression: Enron, Lehman, WorldCom excluded at documented distress dates

Still phase 2b: PFD distress logit, EDGAR fraud wiring, ML fusion.

Acceptance criteria

• Beneish M-Score from multi-period PIT inputs (beneish_v1)
• STA and SNOA from PIT fundamentals (accrual_v1)
• COMBOACCRUAL bottom-5% gate per QVAL process
• Beneish bottom-5% gate per QVAL process
• Exclusions parquet schema matches permanent-loss spec
• CI fails if Enron, Lehman, or WorldCom stop being flagged on fixture dates
• Spec + CONTEXT.md updated; phase 2b backlog trimmed (PFD next)
• MLflow exclusion counts per rule (deferred to Phase 2: QV funnel orchestrator and production scoring #91)

PRs

• feat: forensic evaluator and permanent-loss filter (#89) #118 — forensic evaluator (distress + Beneish)
• TBD — STA/SNOA/COMBOACCRUAL extension

Blocked by

• Phase 2: Multi-period fundamentals ETL and PIT history #87 (closed)

[JLaborda]

This was generated by AI during triage.

Agent Brief

Category: enhancement
Summary: Implement phase-2 forensic evaluator with distress rules, Beneish M-Score, bottom-5% QVAL gate, and CI regression for Enron/Lehman/WorldCom.

Current behavior:
No permanent-loss / forensic module exists. Sector exclusions run upstream in universe construction only. Multi-period PIT fundamentals (#87) are available on develop.

Desired behavior:

• Pure-function distress evaluators (Altman Z, interest coverage, net debt/EBITDA, negative equity, delisted) with versioned YAML thresholds.
• Beneish M-Score from multi-period PIT inputs; bottom 5% cross-sectional hard exclusion on survivors.
• EDGAR-dependent fraud rules log unavailable without blocking.
• Write curated/permanent_loss/run_date=<date>/exclusions.parquet with full schema.
• Missing inputs → review queue, not silent pass.
• CI regression: Enron, Lehman, WorldCom excluded at documented fixture dates.

Key interfaces:

• run_forensic_evaluator(data_dir, run_date) → writes exclusions parquet + optional review queue.
• Rule evaluators return rule_id, rule_version, triggered_value, threshold, explanation.
• Config under config/permanent_loss/ (coefficients/thresholds versioned for future book alignment).

Acceptance criteria:

• Beneish M-Score from multi-period PIT with versioned config
• Bottom-5% forensic percentile gate per QVAL process
• Exclusions parquet schema matches permanent-loss spec
• CI fails if Enron, Lehman, or WorldCom stop being flagged on fixture dates
• Spec 008 updated (implementation status, Beneish promotion)

Out of scope:

• QV funnel wiring (#91)
• run-demo-pipeline integration
• MLflow logging
• Live EDGAR fraud data