The met office raised the warning to signal seven on a Wednesday evening, and by eight o’clock the managing director of Upokul Microfinance was standing in front of a wall map of the coastal belt asking the question every MFI leader in Bangladesh has asked before a landfall: how much of our book is under that cone?

Upokul runs ninety branches across Khulna and Barishal divisions - 180,000 active borrowers, weekly group repayment, average loan around 45,000 taka. Its data was, in the way of the sector, everywhere and nowhere: loans and schedules in the core banking system, bKash repayments settling into a separate store, field officers logging visits on a tablet app, group registers digitised into a fourth database that mostly nobody opened. Head office saw portfolio-at-risk by branch once a month, in a pack compiled by hand.

So the honest answer to the MD’s question, the last time a cyclone came through, had been: in about three weeks. Field officers phone in damage assessments as roads reopen; branch managers total them on paper; the regional offices consolidate; head office finally learns its exposure at roughly the same time the arrears it could have prevented begin hardening into defaults. Three weeks is not a reporting inconvenience in this business. It is the entire window in which a shocked borrower can still be saved.

What the water does to a loan book

Microfinance survives on the tightest repayment discipline in banking, and covariate shock is the one thing that discipline cannot absorb. A flood does not impair a borrower; it impairs a district - every member of every group at once, which is precisely the event the group guarantee was never built for. The numbers, when someone has been positioned to measure them, are brutal in their speed. During the 2018 Kerala floods, an analysis of 27,000 affected microloans found collection efficiency collapsed from 96 percent to 60 percent within the flood month, while the share of portfolio with any missed payment jumped from 0.2 percent to 23 percent. One month, fully formed. The same analysis found recovery was quick on average but sharply uneven across villages and borrowers - which means the average, the only number a monthly branch report can offer, is exactly the wrong instrument for managing it. And the downside when an institution flies blind through such an event is a matter of record: the 1987 floods left roughly a quarter of Grameen Bank’s borrowers in default, and it took a government bailout to steady the institution.

Collection efficiency 96% pre-flood60% flood month

Portfolio at risk (PAR-0) 0.2% pre-flood23% flood month

0%25%50%75%100%

Recovery was fast on average but uneven by village and borrower - invisible without granular, current data. In 1987, floods left roughly 25% of Grameen Bank’s borrowers in default; a government bailout followed.

Sources: Northern Arc portfolio analysis via World Bank blogs (2024); Dowla on Grameen Bank, cited in IJDRR (2018).

The sector knows the playbook that prevents the slide - grace periods for the genuinely affected, fresh recapitalising credit, restructuring before arrears calcify. What it chronically lacks is the targeting. A grace period granted to everyone in a division destroys repayment culture; granted to no one, it converts a natural disaster into a solvency problem. Granted to the right borrowers, within days, it is the difference between a bad quarter and a crisis. The binding constraint has never been willingness. It is that most MFIs cannot produce a current, geo-targeted list of exposed borrowers in under a week, because that list lives across four systems that have never been joined.

Wednesday, 20:40

This time the MD’s question had somewhere to go. Upokul had put the Analytics Engine over all four stores months earlier - read-only connections, every query logged - mostly, at first, for the mundane reporting it will get to below. Standing at the map, the head of risk typed:

> All active borrowers in upazilas under signal seven or higher: outstanding balance, next installment date, savings balance, date of last field visit. Total exposure by branch. Text each field officer their own affected-client list.

By 21:15 head office knew its exposure to the taka, per branch, and eleven hundred field officers’ phones held the names they would need when the wind died. No committee had convened. No spreadsheet had been emailed. The question that used to take three weeks of phone calls had taken thirty-five minutes, most of which was spent deciding the signal threshold.

The storm made landfall Thursday night. What happened over the following six weeks is the part that never makes the disaster coverage. Because the wallet settlements sat in the same query layer as the loan book, Upokul could ask a question that separates distress from disengagement: which affected borrowers have resumed savings deposits but not loan repayments? A family still putting away twenty taka a week has not abandoned its obligations; it is triaging them. Those borrowers - thousands of them - got proactive restructuring offers and grace periods in week two, before a single group meeting had to turn into a confrontation. Borrowers whose activity had stopped entirely got a field visit instead. The recovery itself was tracked at the resolution Kerala proved necessary: collection efficiency by union, weekly, with an alert on any union still below 80 percent after six weeks - because an average that says “recovered” can hide three villages that haven’t, and those three villages are next quarter’s write-offs.

• Core banking systemloans, schedules, PAR
• Mobile-money settlementsbKash / wallet repayments
• Field-officer appvisits, collections, notes
• Digitised group registersmembership, guarantees

Analytics Enginenatural-language query + alert engine

• SMS to field officersaffected-client lists after a shock
• Branch PAR alertsthreshold breaches, same day
• Manager digestsmissed-installment lists, ranked
• Live dashboardsPAR30 by branch / officer / district

Read-only connections; every generated query is logged. No system replacement, no migration.

The unglamorous weeks in between

Cyclones are the demonstration; the standing rules are the product. The one that changed Upokul’s ordinary economics is almost embarrassingly simple: every Friday, each branch manager receives a generated list of borrowers who have missed two consecutive weekly installments, grouped by loan officer and village, ranked by outstanding balance.

The sector’s vital sign, PAR30, is a lagging indicator by construction - by the time a loan crosses into the thirty-day bucket, four weekly installments are already gone and the monthly pack is the first anyone at head office hears of it. Two missed installments is where a field visit still works: the borrower is embarrassed rather than resigned, the group still has leverage, the problem is a cash-flow hiccup and not yet an identity. Upokul’s intervention point moved three weeks earlier per case, in a product whose whole term is often under a year. On top of that sits a portfolio tripwire - an SMS to the regional head if any branch’s PAR30 climbs more than two points inside a fortnight - which has twice caught what turned out to be an officer problem rather than a borrower problem, weeks before the monthly pack would have hinted at it.

There is a quieter dividend that MFI leadership feels most at funding time. Upokul borrows from apex lenders and international debt funds, and every facility renewal used to trigger the same scramble: three weeks assembling portfolio cuts to someone else’s template, followed by due-diligence calls spent defending numbers the team had hand-built under deadline. The last renewal went differently - the lender’s questions were answered in the meeting, from the same governed layer management uses itself, including the union-level recovery curves from the cyclone quarter. Institutions that can show a funder exactly how their book behaved through a disaster, at that resolution, are not many; the pricing conversation noticed. Transparency, it turns out, compounds outward as well as inward.

And the monthly pack itself is gone. The two head-office staff who compiled it now spend their time on the questions the compiling used to crowd out - this quarter, a branch-level analysis of which loan purposes deteriorate fastest after weather shocks, run as a conversation with the data rather than a systems project. Same people, entirely different value.

What an MFI is actually buying

Strip away the software and Upokul acquired three things it had never had. Transparency: the MD’s map-room question - the single most consequential question in coastal microfinance - now has a thirty-five-minute answer instead of a three-week one. Awareness: the portfolio announces its own deterioration, borrower by borrower and union by union, to the person positioned to act, while acting still works. And decisions made informed: grace periods aimed by evidence at the genuinely distressed, renewals and provisioning grounded in recovery curves rather than averages, a board that debates what to do about the numbers rather than whose numbers to believe.

None of it required replacing a single system. The core banking stayed. The bKash pipeline stayed. The tablets stayed. What changed is that the institution can finally see itself - all of itself, at once, at the speed of the weather.