Entity Stress Pipeline

What this pipeline does

The pipeline reads every document in financial_kg.companies, extracts CRISIL rating signals from crisilRatings[] and bankFacilities[], maps them to empirical probability-of-default values from CRISIL's own published tables, and outputs a single composite stress score (0–100) per entity.

The output CSV is sorted highest-stress first, making it immediately usable for watchlist monitoring, knowledge-graph edge weighting, or counterparty risk dashboards.

MongoDB fields consumed

Embedded CRISIL FY2025 tables

All PD and transition values are lifted directly from the CRISIL publication. No third-party estimates are used.

Table 1 — Long-Term 1-Year CDRs (FY15-25, Monthly Static Pools)

Modifier notches (+/−) are linearly interpolated between published anchors. E.g. AA+ = 0.03%, AA− = 0.08%, BBB− = 1.10%, BB+ = 1.90%.

Table 3 — Short-Term 1-Year CDRs (FY15-25)

Table 2 — 1-Year Downgrade Probability

Derived from the transition matrix as: 1 − stability_rate − upgrade_rate

Outlook Multipliers

How the stress score is computed

Four Weighted Components

Normalisation

Each component is mapped to [0, 1] before weighting:

• C1, C2 (PDs): Log-linear — ln(1 + 99·PD) / ln(100). AAA→0.0, D→1.0. Log scale captures the non-linear nature of credit risk.
• C3 (Downgrade prob): Linear — min(dp / 0.25, 1.0). Anchored at ~25% max (C-rated entity).
• C4 (Heterogeneity): Linear — min(stdev / 0.50, 1.0). Max theoretical stdev across AAA–D mix ≈ 0.50.

Composite Formula

Stress Labels

Risk Tier Classification

entity_stress_scores.csv — full column reference

One row per entity, sorted descending by stressScore.

Identity

Rating & PD

Facility Exposure

Score Components & Final Output

Running the pipeline

Prerequisites

pip install pymongo

Basic run (all documents)

python entity_stress_pipeline.py \
    --uri  "mongodb://localhost:27017" \
    --db   financial_kg \
    --col  companies \
    --out  entity_stress_scores.csv

Test run (10 documents)

python entity_stress_pipeline.py --limit 10 --out test_output.csv

MongoDB Atlas

python entity_stress_pipeline.py \
    --uri "mongodb+srv://user:[email protected]/"

CLI Arguments

Console Output Example

# After processing completes:
✅  1247 entities written → entity_stress_scores.csv  (errors: 0)

Stress Label Distribution:
  Minimal                  312  ████████████
  Low                      418  ████████████████
  Moderate                 289  ███████████
  Elevated                 142  █████
  High                      63  ██
  Severe                    23  █

Stress Score  min=0.0  max=68.4  mean=14.2  median=11.8

Indian Oil Corporation Limited (INDOIL)

IOC holds Crisil AAA/Stable on both its Term Loan and Short-Term Loan, and all 36 bank facilities are rated AAA/Stable.

Step 1 — Parse crisilRatings[]

• Primary LT = AAA → PD = 0.00%
• Primary ST = A1+ → PD = 0.02%
• Outlook = Stable → multiplier = 1.00×
• Adjusted LT PD = 0.00% × 1.00 = 0.00%
• 1-yr downgrade probability (AAA) = 0.00%

Step 2 — Exposure-Weighted PD from bankFacilities[]

All 36 facilities rated AAA → every facility PD = 0.00%.

• EW-PD = 0.00% | Adjusted EW-PD = 0.00%
• Rating heterogeneity (std dev) = 0.0000 — fully uniform

Step 3 — Normalise & Composite

Design decisions & known limitations

Why log-linear normalisation for PD?

Credit risk is non-linear. The economic gap between BBB (0.46%) and B (8.40%) is far larger than the raw arithmetic difference suggests. A log transform compresses the upper tail and spreads the lower range, giving better resolution to investment-grade distinctions while capturing sub-IG severity accurately.

Why 35 / 30 / 20 / 15 weights?

The primary LT rating (C1) is CRISIL's authoritative credit opinion, so it anchors the score. Facility exposure weighting (C2) captures lender-level concentration risk. Downgrade probability (C3) adds a forward-looking dimension. Heterogeneity (C4) is a secondary signal for internal rating inconsistency across facilities.

Known Limitations

Extending the Pipeline

1. Financial ratio stress — multiply stressScore by a leverage factor derived from mcaPaidupCapital / mcaAuthorizedCapital.
2. Industry adjustment — high-default industries (CRISIL Annexure 7) carry an industry beta multiplier.
3. Time-to-default overlay — combine with CRISIL Annexure 8 survival curves for 2-year and 3-year variants.
4. Graph integration — feed stressScore as edge weights into the financial_kg knowledge graph for counterparty contagion analysis.

Sources

• CRISIL Ratings Annual Default and Ratings Transition Study — FY2025 (Tables 1, 2, 3, 4)
• CRISIL Default Study FY2025 — Annexure 10: Methodology (monthly static pools, marginal default rate method)
• CRISIL Default Study FY2025 — Annexure 7: Industry-wise classification of defaults
• CRISIL Default Study FY2025 — Annexure 8: Analysis of defaults — Time to default
• financial_kg.companies MongoDB collection schema
• pymongo Documentation — pymongo.readthedocs.io