Rvs credit scoring

Rescue Viability Score (RVS)

A proprietary 6-variable logistic regression model with a hazard survival layer for rescue-finance credit decisions. Validated at 93.0% forward accuracy (hazard AUC 0.985, leave-one-out AUC 0.979, mean calibration error 0.029, Brier skill score 0.793) on a 47-company distress cohort. Outperforms the Altman Z-Score by approximately 30 percentage points on distressed rescue targets.

Architecture

Financial Statements → 8 Raw Numbers → 6 Variables (V1–V6)
  → Logistic Regression P(Recovery)
  → Discrete-Time Hazard Layer → 5-Year Survival S(5)
  → Composite Score → Verdict (GO / CONDITIONAL / NO-GO)

Variables

Variable	Formula	Signal
V1	Working Capital / Total Assets	Liquidity
V2	Retained Earnings / Total Assets	Cumulative profitability
V3	EBITDA / Total Debt	Debt coverage
V4	Operating Cash Flow / Total Debt	Cash flow coverage
V5	Net PPE / Total Liabilities	Collateral backing
V6	Revenue / Total Assets	Asset turnover (negative coefficient — captures "high turnover, thin margin" distress)

v3.0 Coefficients (Listed Companies)

Logistic regression: Intercept +2.5445 · V1 +0.2506 · V2 +1.7070 · V3 +0.7426 · V4 +0.7262 · V5 +0.8278 · V6 −1.8122

Hazard model: Constant −1.9295 · V1 −0.2312 · V2 +0.3298 · V3 +0.7437 · V4 −7.1443 · V5 −27.7134 · V6 +3.8640 · TimeDummy(Yr 4–5) −0.8840 · MacroShock(2020–21) +3.8256

Composite: V3_score = 0.60 × P(Recovery) + 0.40 × S(5yr)

Verdicts: GO ≥ 0.70 · CONDITIONAL 0.50–0.69 · NO-GO < 0.50

Mandatory stress tests (4): Working Capital Crisis (V1 × 0.50), Margin Compression (V3 × 0.60), Refinancing Crisis (V4 × 0.40), Perfect Storm (all combined).

v4.0 Extensions (Private Companies)

Analyst adjustments before variable computation: owner-compensation add-back, related-party-transaction EBITDA haircut, collateral appraisal factor (0.70–1.00), revenue under-reporting factor (1.00–1.10).

Hazard logit: z_h(t) = −3.0 + 0.15V1 + 0.80V2 + 0.45V3 + 0.40V4 + 0.50V5 − 0.90V6 + governance_penalty + concentration_penalty, with 5% annual variable decay (V(k,t) = V(k,0) × 0.95^(t−1)), a 1.20× private-company hazard multiplier, governance penalty −0.005 × (100 − GovernanceScore) and concentration penalty +0.003 × ConcentrationScore.

Composite: V4_score = P(Recovery) × S(5yr) × (1 − InfoAsymmetry) × GovernanceFactor, where GovernanceFactor = 0.80 + 0.20 × (GovernanceScore / 100).

Verdicts: GO ≥ 0.65 · CONDITIONAL 0.50–0.64 · NO-GO < 0.50. Sharia non-compliance (AAOIFI screens: Debt/Assets < 33%, (Cash + interest-bearing securities)/Assets < 33%, non-compliant revenue < 5%) applies an automatic one-tier verdict downgrade.

Mandatory stress tests (6): the four v3.0 scenarios plus Key-Person Loss and Governance Collapse. Stress resilience rating: STABLE (0 downgrades) / MODERATE (1–2) / FRAGILE (3+).

Seven Critical Analyst Errors (enforce in every output)

1. Including trade payables in Total Debt — use interest-bearing borrowings only
2. Using gross PPE instead of net PPE for V5
3. Zeroing V2 when Retained Earnings are negative — use the actual negative number
4. Being alarmed by V6's negative coefficient — it is intentional
5. Mixing fiscal years across the 8 raw numbers
6. Setting the macro-shock dummy to 1 for current analysis — it is 1 only when backtesting 2020–2021
7. Forgetting to cap V3/V4 at 2.0 when Total Debt = 0

Output Protocol

Present the 8 raw numbers in an extraction table, show all 6 variable calculations, display the Z-score term by term, convert to P(Recovery), compute hazard rates year by year, calculate the composite with the formula shown, state the verdict with the threshold it cleared, run all stress tests in a summary table, and state the stress resilience rating. Show every coefficient and intermediate calculation — the output must be defensible in an investment committee.

Resources

• knowledge/methodology.md — full model methodology and validation results
• dataset/train.jsonl — instruction-tuning pairs
• eval/eval.jsonl — held-out evaluation questions