EMI detection starts with text signals: EMI, loan repayment, home/car/personal loan, and lender names (HDFC, Bajaj, Tata Capital, and others). But keywords alone over-count — one-off transfers mention 'loan' too.
The fix is recurrence: group keyword-matched debits by rounded amount and keep groups seen 2+ times as confirmed EMIs. Single explicit mentions are still reported, flagged unconfirmed, so analysts can review instead of missing them.
A useful response includes the count, total paid in period, estimated monthly EMI, and a per-lender breakdown with totals and counts — plus the raw samples so your team can audit.
Combine monthly EMI with income to get debt-to-income (monthly EMI / monthly income) and fixed-obligations ratios. Those two numbers drive more credit decisions than any other pair in the response.