Credit Risk Intelligence (Streamlit)

Live demo: https://credit-risk-intel-auro.streamlit.app

Repo: https://github.com/Aurokrishnaa/credit-risk-intelligence-Auro

Stress, assess, and price credit portfolios with explainable PD/LGD/EAD analytics, CECL/IFRS lifetime loss, covenant headroom, and a print-ready executive report, built in Python/Streamlit by Aurokrishnaa R L (MS Finance—Quant · MBA Finance).

---

Table of contents

• Highlights
• What’s inside (end-to-end flow)
• Screenshots
• Quickstart
• Usage guide
• Data schema
• Modeling details
• Risk tiers & stress design
• Covenants & watchlist
• CECL / IFRS
• Risk-based pricing
• Report Mode (print/PDF)
• Performance & limits
• Project structure
• Troubleshooting
• Roadmap
• License
• Author & contact

---

Highlights

• Robust data ingestion: CSV template (download in-app), tolerant encoding/sep loader (UTF-8/UTF-8-SIG/Latin-1; comma/semicolon) and clean rate parsing.
• Core risk stack: PD / LGD / EAD / EL at loan level with transparent heuristics and safeguards.
• Macro stress testing: rate (bps), unemployment (pp), and collateral shocks flow through cashflow metrics → PD/LGD/EAD/EL.
• Covenants & watchlist: DSCR, ICR, LTV, DTI headroom with Green/Amber/Red badges and a downloadable exceptions file.
• CECL/IFRS lifetime EL: amortization curve + monthly default hazard + PV discount.
• Risk-based pricing: EL% + (capital + opex + profit) → suggested spread (bps) and all-in rate.
• Executive reporting: polished Report Mode with snapshot locking, PII masking, charts, and print/PDF styles.
• Tech: Streamlit · pandas · NumPy · numpy-financial · Matplotlib.
• App version: v1.3.0.

---

What’s inside (end-to-end flow)

Data → Base risk → Stress → CECL → Covenants → Pricing → Report

1. Ingest & validate: upload CSV or click Use sample dataset.
2. Derive cashflows & ratios: payment, balance @ 12m, interest Y1, LTV/DTI/DSCR/ICR.
3. Estimate base risk: PD (scorecard), LGD (collateral haircuts), EAD (12m avg), EL.
4. Apply stress: rate/unemp/collateral shocks propagate; PD/LGD/EAD/EL recomputed.
5. Tier & concentrate: PD-percentile tiers (Low/Med/High); top sectors/geos by EAD.
6. Covenants: compute headroom, badge (R/A/G), watchlist & export.
7. CECL/IFRS: lifetime EL (PV) using monthly default intensity over amortization.
8. Pricing: translate risk into suggested spread (bps) and all-in rate.
9. Report Mode: one-click print/PDF with KPI deltas, composition, watchlist, and pricing insights.

---

Screenshots

---

Quickstart

1) Run the live app (no setup)

• https://credit-risk-intel-auro.streamlit.app

2) Run locally

# clone
git clone https://github.com/Aurokrishnaa/credit-risk-intelligence-Auro.git
cd credit-risk-intelligence-Auro

# (optional) create a venv
python3.12 -m venv .venv
# mac/linux
source .venv/bin/activate
# windows (powershell)
# .venv\Scripts\Activate.ps1

# install
python -m pip install -U pip
pip install -r requirements.txt

# run
streamlit run app.py

Requirements

• Python 3.12
• streamlit, pandas, numpy, numpy-financial, matplotlib (pinned in requirements.txt)

---

Usage guide

1. Get started

   • Click Download CSV template, replace sample rows, keep the same headers.
   • Or click Use sample dataset to explore features instantly.
   • Upload your CSV (≤ 200 MB).

2. Scenario & filters

   • Sidebar: choose Base / Adverse / Severe / Custom (rate bps, unemployment pp, collateral drop %).
   • Global filters: Sector, Geography, and Risk tier (stressed).

3. Explore tabs

   • Overview: KPIs (EAD / w-avg PD / EL / High-risk%) and key charts.
   • Data: first rows + derived metrics; validation issues (if any).
   • Base Risk: loan-level PD/LGD/EAD/EL and PD-tiering.
   • Stress Test: shocked PD/LGD/EAD/EL and tier mix.
   • Covenants & Watchlist: DSCR/ICR/LTV/DTI headroom, badges, exceptions CSV.
   • CECL / IFRS: lifetime EL (PV) with discount rate.
   • Pricing: suggested spread (bps) & all-in rate; export pricing table (CSV).
   • Reports: Report Mode → Print / Save as PDF (browser dialog).

4. Notes

   • InterestRate is numeric. For 7.5%, use 0.075 or 7.5 (no % sign).
   • Minimal required columns: LoanAmount, InterestRate, TermMonths.

---

Data schema

Minimal required

• LoanAmount (USD)
• InterestRate (decimal or percent numeric; e.g., 0.075 or 7.5 for 7.5%)
• TermMonths (>0)

Recommended (for richer analytics)

• LoanID, Borrower, Sector, Geography
• CreditScore (300–850)
• MonthlyIncome, NOI, EBITDA
• CollateralType, CollateralValue, CurrentMarketValue
• ExistingDTI

The app auto-selects CurrentMarketValue (if valid) else CollateralValue as CollateralBase.

Validation checks

• Positive LoanAmount / TermMonths, valid numeric InterestRate
• CreditScore in [300, 850] (if provided)
• Collateral presence/positivity
• Rows with issues appear in Data → Rows with data issues

---

Modeling details

Cashflow & ratios

• Monthly payment A = pmt(r/12, n, -P) with zero-rate guard.

• Outstanding balance @ 12m: amortization formula with edge-case handling.

• First-year interest: month-by-month loop.

• Ratios:

  • LTV = LoanAmount / CollateralBase
  • DTI = MonthlyPayment / MonthlyIncome
  • DSCR = NOI / AnnualDebtService (or MonthlyIncome×12 / AnnualDebtService as fallback)
  • ICR = EBITDA / Interest_Y1

PD (12-month) — scorecard heuristic

• Base by CreditScore bands (prime → subprime), then multiplicative factors:

  • DTI buckets (↑ leverage → ↑ PD)
  • LTV buckets (↑ leverage → ↑ PD)
  • Sector factor (e.g., hospitality/retail > manufacturing/healthcare)

• Clamped to [0.2%, 35%].

LGD — collateral haircuts

• Haircuts by CollateralType (e.g., real estate 30%, equipment 45%, inventory/AR 60%, securities 10%).
• Recovery cost (10%) then LGD = 1 − min(effective collateral / LoanAmount, 1)×(1 − recovery_cost); clamp [5%, 95%].

EAD (12m)

• EAD_12m = 0.5 × (LoanAmount + Balance@12m).

EL (12m)

• EL_12m = PD_12m × LGD × EAD_12m.

---

Risk tiers & stress design

Tiers (explainable)

• Percentile-based on base PD (≈33/66 cuts) with a stable fallback when samples are small.
• Stressed tiers reuse base PD cutpoints for comparability.

Stress scenario mechanics

• Rate shock (bps) → raises payment/DSCR/DTI → feeds PD & balances.
• Unemployment shock (pp) → multiplicative PD uplift min(1 + 0.10×pp, 2.5).
• Collateral drop (%) → lower CollateralBase → higher LTV → feeds LGD/PD.
• All recalculated metrics (PD/LGD/EAD/EL) are shown under Stress Test.

---

Covenants & watchlist

• Headroom under stress:

  • HR_DSCR = DSCR_st − DSCR_min
  • HR_ICR = ICR_st − ICR_min
  • HR_LTV = LTV_max − LTV_st
  • HR_DTI = DTI_max − DTI_st

• Badges: Red (breach), Amber (near), Green (healthy) using user thresholds + amber buffers.

• Watchlist: Red/Amber sorted by WorstHeadroom; Download Exceptions (CSV).

---

CECL / IFRS

• Monthly default probability from 12-month PD: p_m = 1 − (1 − PD_12m)^(1/12).

• Lifetime EL (PV) loops over months:

  • survival × p_m gives default mass each month
  • EAD_t is outstanding balance
  • EL_t = survival × p_m × LGD × EAD_t
  • discount by (1 + r/12)^t (user-set discount rate)

• Reported as totals and as % of start EAD for Base and Stressed.

---

Risk-based pricing

• EL rate (stressed): EL_12m_st / EAD_12m_st.
• Suggested spread: EL% + (capital buffer + opex + target profit).
• All-in rate: funding cost + spread.
• Portfolio-level metrics + downloadable pricing table (CSV).

---

Report Mode (print/PDF)

• Snapshot lock to freeze metrics and charts for printing.
• Mask sensitive IDs (LoanID/Borrower) toggle for sharing.
• Detailed View explains methods and stack inline.
• Browser print dialog → Print / Save as PDF (letter/A4, default/narrow margins, hide headers/footers).

---

Performance & limits

• Upload limit: 200 MB (CSV).
• Caching: heavy calcs wrapped with @st.cache_data (payments, balances, lifetime EL).
• Matplotlib runs on the Agg backend for reliable image generation.

---

Project structure

credit-risk-intelligence-Auro/
├─ .streamlit/              # Streamlit config/theme
├─ data/                    # Template/sample CSVs (written by the app)
├─ app.py                   # Main Streamlit application
├─ requirements.txt         # Python dependencies
├─ LICENSE                  # MIT
└─ Assets/                  # Screenshots of the app

The app may create reports/ at runtime for convenience; it is not required to be committed.

---

Troubleshooting

• “Unsupported or unreadable CSV” → export as UTF-8, comma-delimited; start from the in-app template.
• “Required columns missing” → ensure headers match exactly. Minimal: LoanAmount, InterestRate, TermMonths.
• Interest rate parsing → use 0.075 or 7.5 for 7.5% (no %).
• Empty view after filters → set Sector/Geography/Tier back to All (the app auto-relaxes the tightest filter and warns).

---

Roadmap

• Odds/Logit PD option with simple calibration helper
• Multi-period EAD (amortizing/utilization profiles)
• Scenario manager (save & compare)
• CSV schema lint with column type hints

---

License

This project is licensed under the MIT License. See LICENSE.

---

Author & contact

Aurokrishnaa R L — MS Finance (Quant) · MBA Finance

• Website: https://www.aurokrishnaa.me
• LinkedIn: https://www.linkedin.com/in/aurokrishnaa

Models are demo-calibrated for illustration only and are not a production credit policy or financial advice.