AI Credit Intelligence System

A full-lifecycle credit-risk study: originate, price, provision, and collect, built on four real public datasets using standard model-risk-governance conventions (SR 11-7, RBI MRM, Basel III, IFRS 9). Each model is measured against a simple baseline, a bureau-score cutoff and its equivalents, the kind of benchmark a lender relies on before it builds a full custom scorecard.

The deliverable is not a higher Gini. It is a better business frontier: at the benchmark's operating point, approve more good borrowers and fewer bad ones, and reduce credit losses across the book's life.

Scope note: the "Champion" throughout is a bureau-score benchmark, not a mature lender's production scorecard. The results show the lift a custom model adds over a generic bureau score. They are not a claim to beat an established underwriting process. See Scope and honest limitations.

---

Headline result

On a held-out test set of 61,503 loans, a custom PD model is benchmarked against a bureau-score cutoff, the baseline a lender uses before a full scorecard exists.

	Bureau-score benchmark (Champion)	Custom (Challenger)
Discrimination (AUC)	0.718	0.756
At matched volume	3.69% default	2.95% default (-20%)
At matched risk	50% approval	63.6% approval (+27%)

Recommended operating point (SC3): 61.9% approval, 3.61% default, +₹176 cr net income, +93.4% portfolio RAROC - higher approval and lower default than the bureau-score benchmark, simultaneously. The income-maximising point within risk appetite sits at 63.7%; the recommended point is held one step inside it to strictly beat the benchmark on every metric.

---

System architecture

flowchart LR
  subgraph Data
    D1[Home Credit]
    D2[Give Me Some Credit]
    D3[LendingClub]
    D4[Freddie Mac panel]
  end
  subgraph Modules
    A[A - Application PD]
    B[B - Behavioural]
    C[C - Portfolio and Pricing]
    E[D - IFRS 9 ECL and Staging]
  end
  G[Decision Engine - RAROC gate]
  P[Provisioning - IFRS 9 ECL]
  GV[Governance - SR 11-7, RBI MRM, Basel III]

  D1 --> A
  D2 --> B
  D3 --> C
  D4 --> E
  A --> G
  B --> G
  C --> G
  G --> P
  E --> P
  P --> GV

Loading

The PD model estimates risk; the RAROC gate sets policy on the frontier; the IFRS 9 engine turns staging into a provisioning decision; governance closes the loop with monitoring and regulatory alignment. The four modules are trained on four independent datasets; the decision engine that combines their signals is a reference architecture, not a single scored book (see limitations).

---

The four risk modules

Each module measures a custom Challenger against a simple benchmark, the kind of baseline a lender uses before building a custom model, and reports the lift on discrimination and on the business outcome.

Module	Question	Benchmark (Champion)	Custom	Lift vs benchmark
A Application	Who to approve?	Bureau score (0.718)	Full PD model (0.756)	+27% volume / -20% default
B Behavioural	How is the borrower behaving?	Past-DPD rule (0.768)	Behavioural ML (0.860)	catches 81% vs 71% of delinquents
C Portfolio	Is the rate right?	FICO + grade (0.705)	Full model (0.729)	+3.7pp approval / -4.9% default
D ECL & Staging	What is the lifetime loss?	Snapshot proxy	Panel staging on 12.6M loan-months	39x Stage 1 to Stage 2 provisioning cliff, real LGD 41.8%

Module D's transition matrix (28.5% monthly Stage 2 to Stage 1 cure rate) and vintage curves (2008 crisis 6.0% vs 2011 1.1% default) drive the IFRS 9 ECL-reduction case: better origination keeps loans in Stage 1, and collections cure Stage 2 back before the lifetime-ECL cliff.

---

Repository structure

01_module_a_application_risk/        Module A - origination PD, scorecard, RAROC, swap-set
02_module_b_behavioural_risk/        Module B - delinquency model, collections capture
03_module_c_portfolio_pricing_risk/  Module C - grade / market-implied PD, pricing
04_module_d_ecl_ifrs9/               Module D - IFRS 9 staging, transition matrix, lifetime ECL
04_decision_engine/                  Composite signal, RAROC gate, audit trail
05_governance/                       Model cards, SR 11-7, RBI MRM, regulatory alignment
06_docs/                             Executive deliverables and methodology (HTML)

Each module follows the same layout: 01_data/ (raw is gitignored, download separately), 02_notebooks/, 03_models/, 04_outputs/.

---

Run it

The decision engine runs with no data download:

pip install -r requirements.txt
python 04_decision_engine/demo.py     # scores 4 sample applicants end to end

To regenerate every figure from raw data, place the datasets (see Data) and run:

python run_all.py            # runs the module pipelines in order, on a fixed seed
python run_all.py --check    # lists which raw datasets are present / missing first
python tools/run_notebooks.py  # re-execute notebooks whose data is present (refreshes outputs)

run_all.py skips any module whose raw data is absent and tells you where to get it, so a partial checkout still runs what it can. Every figure in the documentation is regenerated by these pipelines on a fixed seed (42).

---

Scope and honest limitations

This is a portfolio study on public data, built to institutional conventions. Read the results with these caveats, which the analysis itself states rather than hides:

• The Champion is a bureau-score benchmark, not a production scorecard. The lift shown is over a generic bureau-score cutoff, the baseline before a custom model exists. It is not a claim to beat a mature lender's existing underwriting.
• The integration is illustrative. The four modules train on four unrelated public datasets (different borrowers, geographies, eras). The composite PD and decision engine show how the signals would combine; they are a reference architecture, not one book scored end to end.
• Module A's origination "staging" is a PD-bucket proxy, not SICR-based IFRS 9 staging. A single application snapshot cannot express a significant increase in credit risk. All credible IFRS 9 staging in this project is Module D, on the Freddie Mac monthly panel.
• The collections uplift / next-best-action model is exploratory. It uses loan modification as an observed (not randomised) treatment, so the effect is confounded, and its uplift-decile validation is not yet monotonic. Treat it as a targeting hypothesis, not a causal result.
• Data is public Kaggle / Freddie Mac. Figures are realistic and reproducible, but they describe these datasets, not a live portfolio.

---

Governance and standards

Built to the following conventions (not a formal validation):

• SR 11-7 - model development, validation, monitoring; champion vs challenger; audit trail.
• RBI MRM - model inventory, risk-tier classification, model cards per module.
• Basel III / ICAAP - EL, RWA, economic capital, four-scenario stress testing.
• IFRS 9 - Stage 1 / 2 / 3 allocation, lifetime PD from the survival curve, ECL by stage.
• Fair Practices - adverse-action notices with plain-language reason codes.

Data

Home Credit Default Risk · Give Me Some Credit · LendingClub 2007-2018 · Freddie Mac Single-Family Loan Performance (2008-2011). Raw files are downloaded separately and are not committed.

---

A portfolio study in credit risk: reduce risk, increase real profitability. Every figure traces to an executed pipeline and is built to standard risk conventions.