हे पान सध्या फक्त इंग्रजीत उपलब्ध आहे.
How it works
An interpretable model, so every reason and every plan is exact.
Pathway pairs a logistic regression credit model with a search over realistic changes, a month-by-month simulator, illustrative risk-based pricing and a Monte Carlo certainty band. Everything below is read from the same code and configuration the app runs.
Results
Headline metrics
Model AUC, validation
0.857
Model AUC, test
0.855
Rejected applicants with a plan
98.4%
Median months to approval
12
Recourse effort gap
10.3%
Data: Kaggle 'Give Me Some Credit' (cs-training.csv, 150,000 rows): 60% training, 20% validation, 20% test, stratified. The cut-off was chosen on the validation rows; the test rows were used once, to confirm it. At the cut-off the model approves 86.6% of validation applicants and catches 59.8% of defaulters (test: 86.6% and 60.3%). Plan figures: all 4,035 rejected test applicants, run through the same TypeScript engine the app ships; every plan counted was re-scored by the model and approved.
Pipeline
From training data to a plan
Step 1: Train offline
Python cleans Kaggle's Give Me Some Credit data, splits it 60/20/20, fits logistic regression on the training rows and picks the cut-off on the validation rows.
Step 2: Export
Coefficients, scaler, intercept, cut-off and cleaning rules go to lib/model.json and lib/model.meta.json. No Python runs in production. Rupee incomes are divided by 20, close to India's purchasing-power parity, on the way into the model.
Step 3: Recourse engine
TypeScript searches every feasible change set, scores each one with the same cleaning and model, and only returns a plan the model itself approves.
Step 4: Timeline
A month-by-month simulator moves each change at a capped pace and finds the approval month.
Model
Features, coefficients and assumptions
Features and what a plan may do with them
| Feature | Class | Coefficient |
|---|---|---|
| Credit card utilization | actionable | +0.635 |
| Payments 30-59 days late | slow-moving | +0.349 |
| Payments 60-89 days late | slow-moving | +0.204 |
| Payments 90+ days late | slow-moving | +0.327 |
| Late-payment history reported as a special code | derived | +0.106 |
| Monthly income | slow-moving | -0.024 |
| Debt-to-income ratio | actionable | +0.162 |
| Open credit lines and loans | actionable | -0.054 |
| Income not provided | derived | -0.048 |
| Income given as 0 or 1 | derived | -0.075 |
Standardized coefficients on the log-odds of serious delinquency; positive means riskier. Income and the three late-payment counts enter as log(1 + value). The three derived features are 0/1 flags worked out from the inputs (a late-payment count reported as code 96/98, no income given, income given as 0 or 1); for an applicant who enters a real income and real counts they are all 0. Age, dependents and real-estate loans are not model inputs.
Assumptions (lib/config.ts)
- Plan horizon
- 36 months maximum
- Card paydown pace
- up to 6% points of card utilisation per month
- EMI reduction
- 5% of today's EMIs per month, 50% max
- Income growth
- 0.6% per month, capped at +15% in total
- Active loans and cards
- change by at most 2, effective after 1 month
- Late payments
- counted over 24 months, assumed evenly spread, age out if every future payment is on time
- Not used by the model
- age, dependents, real-estate loans
Approval cut-off: a predicted default probability below 10% is approved; exactly 10% or above is declined. 10% sits at Pathway score 650. The cut-off was chosen on the validation rows, where it declines 13.4% of applicants. Every 50 points doubles the odds of repaying.
See the fairness auditMoney saved
Money saved: risk-based pricing
IllustrativeUnder risk-based pricing, a lower-risk borrower is offered a lower interest rate. Pathway turns a plan's score gain into money by pricing the same loan twice: at today's score and at the score after the plan. The tiers below were chosen for this demo. They are illustrative, not any lender's pricing, and not an offer of credit.
Tiers on the default loan
| Tier | Pathway score | APR | EMI | Total interest |
|---|---|---|---|---|
| Excellent | 760+ | 10.5% | ₹6,500/mo | ₹34,018 |
| Very good | 720+ | 12.5% | ₹6,691/mo | ₹40,866 |
| Good | 690+ | 15% | ₹6,933/mo | ₹49,590 |
| Fair | 650+ | 18% | ₹7,230/mo | ₹60,297 |
| Declined: high-cost alternative | below 650 | 36% | ₹9,161/mo | ₹1,29,787 |
Pricing assumptions (lib/pricing.ts)
- Score 760+
- 10.5% a year
- Score 720+
- 12.5% a year
- Score 690+
- 15% a year
- Score 650+
- 18% a year
- Below 650 (declined)
- 36% a year from a high-cost lender
- Default loan
- ₹2,00,000 over 36 months
- Affordability
- FOIR: all EMIs at most 50% of monthly income
Formula
EMI equals P times r, divided by 1 minus (1 plus r) to the power of minus n.
P is the amount borrowed, r the monthly rate (yearly rate ÷ 12) and n the number of monthly payments. Interest saved is total interest at today's rate minus total interest at the after-plan rate, on the same loan.
Worked example
Crossing the approval line, from declined (36% a year from a high-cost lender) to the Fair tier (18% a year), cuts projected interest on the default loan from ₹1,29,787 to ₹60,297:
₹69,490 saved
and a monthly payment ₹1,930 lower. Illustrative.
Uncertainty
How sure is the timeline? Monte Carlo
SimulationMany futures, not one
The month-by-month projection moves every change at its capped pace: one tidy future. Real life is noisier, so Pathway also replays the same plan as 400 simulated futures.
Each future draws its own pace of card paydown and debt reduction and its own income growth (still capped by the plan's target), and may be hit by an income shock that pauses progress or by a new late payment. The random draws are seeded, so the same applicant always gets the same band and the result is repeatable.
From the approval month of every future, Pathway reports the 10th percentile as the best case, the median as the likely month and the 90th percentile as the worst case. The shaded band around the projection spans the 10th to 90th percentile of the simulated futures each month. A future that has not reached approval by month 36 counts as “beyond 36 months”, and the share of futures approved within the horizon is shown next to the band.
It is a simulation of plausible paths, not a forecast for any individual, and the settings are illustrative.
Simulation settings (lib/config.ts)
- Simulated futures
- 400 per applicant (seeded, repeatable)
- Pace of paydown
- 40%–130% of the plan's pace, centred on 100%
- Income growth
- 0.0%–1.5% per month, centred on 0.6%
- Income shocks
- 1.5% chance a month, pausing progress 2–3 months
- New late payment
- 0.6% chance a month
- Band shown
- 10th to 90th percentile; "likely" is the median
Goal planner
Instead of asking “what gets me approved?”, the goal planner starts from the rate you would like. It works backwards from a target yearly interest rate to the lowest Pathway score whose tier is priced at or below it, then runs the same lowest-effort search the approval plan uses, aimed at that score instead of the cut-off.
For example, a target of 15% a year needs a score of at least 690, the Good tier in the illustrative pricing above.
It also applies an affordability rule: all monthly instalments together, including the new loan, should stay within 50% of monthly income (PRICING.maxEmiToIncome). Indian lenders call this the FOIR limit. When the payment does not fit, the planner also works out the largest amount, or the shortest longer term, that would.
Offer check
Enter the terms of a loan offer and the offer check works out its true cost from the cash flows: the money you actually receive after any upfront fees, against every repayment you make.
The true APR is the internal rate of return (IRR) of those cash flows, solved numerically, and annualised with a simple day-count convention: the daily rate × 365, with monthly instalments counted as 30 days. That puts offers with different repayment schedules on one scale. The effective annual rate, (1 + daily rate)365 − 1, is shown alongside.
Red-flag thresholds are illustrative. A flag is a prompt to read the terms and ask questions, not a verdict on any lender. The calculation runs entirely in your browser.
At the same risk, who has to work harder?
The fairness audit compares recourse effort across age and income bands, adjusted for risk.