Build a simple delinquency model for a mortgage pool using monthly roll rates between arrears buckets, project the 90+ bucket for three months, and say which borrower factors you would add.

The 90+ bucket grows from Rs 80 crore to about Rs 140, 185 and 228 crore over three months, from 0.8% to about 2.3% of the pool. Each month about 3% x 30% x 50% of current loans, roughly Rs 43 crore, completes the journey. Borrower factors such as loan-to-value, income cover, credit score and employment type turn the single roll rates into loan-level probabilities.

Think of a queue of students who have missed homework. Most who miss one hand it in next week; a few miss two; some of those miss three and are sent to the principal. A roll-rate model tracks how much of each arrears bucket moves one step worse each month, so today's early misses become a forecast of tomorrow's defaults. Loans move at most one bucket a month. With a roll rateThe share of loans in one arrears bucket that move to the next, worse bucket in the following month. of 3% from current, 30% from 30 days and 50% from 60 days, only 0.45% of current loans reach 90+ three months later, but on Rs 9,500 crore that is about Rs 43 crore every month.

The early buckets are the forecast: today's Rs 300 crore at 30 days and Rs 120 crore at 60 days already determine most of the next two months of new defaults. That is why investors watch 30-day arrears closely; they move first. The limits are worth saying. Roll rates are averages that change with the economy, rising when rates reset upward or jobs are lost. One rate for the whole pool treats a new loan at 90% loan-to-value the same as a ten-year-old loan at 40%. And the model ignores cures from 90+, write-offs, prepayments and new lending, each of which changes the balances.

Add the factors that explain why one borrower misses a payment and another does not. Loan-to-value, because a borrower with equity in the home fights to keep paying; the share of income taken by loan payments; credit score and past arrears; employment type, salaried against self-employed; and loan age, because missed payments tend to cluster in the early years. Add rate resets for floating loans, geography, and property type. To build it, take loan-level monthly history, estimate the probability of each transition as a function of those factors, for example with a logistic regression per bucket, and validate on a later period the model did not see. Then run it forward under stress: raise the current-to-30 rate from 3% to 5% and watch what happens to the 90+ bucket in month 3.