Step 02 · Prediction Engine
In progressUpokul Microfinance is the same composite as in the Analytics Engine cases. The scenario is illustrative; the industry figures are sourced. The Prediction Engine is in active development; this case describes the workflow it is being built to run.
The Friday list changed Upokul’s economics. Every branch manager receives the borrowers who have missed two consecutive weekly installments, grouped by loan officer and village, and a field visit follows while it still works - three weeks earlier than the portfolio-at-risk report would ever have flagged the loan. The head of risk’s question, a year later, was the obvious one. If two missed installments is early, what does earlier look like?
Field officers could describe it from experience. A borrower who used to pay by bKash two days before the group meeting starts paying cash on the day, at the last minute. Weekly savings deposits shrink, then stop. Another member of the same group falls behind first. A loan taken for an agricultural purpose enters the lean weeks before harvest. None of those is a missed installment. All of them are visible in data Upokul already holds.
The regulatory frame makes the timing unforgiving. Bangladesh’s microfinance institutions serve nearly 34 million borrowers, and classify a loan as non-performing once it is more than a month overdue - a far stricter standard than the banking sector’s. On a weekly repayment schedule, a month is four installments. By the time a loan is non-performing, the work has shifted from prevention to recovery.
The research case for predicting earlier is well established. Using behavioural patterns in mobile phone records, Björkegren and Grissen showed that borrowers in the highest predicted-risk quintile were 2.8 times as likely to default as those in the lowest, and the result held when tested on a later time period. Upokul does not use call records and would not. But its own repayment timing, savings behaviour and group records are behavioural data it already holds as part of the lending relationship.
The traditional route to a model like that, for an institution of Upokul’s size, is a donor-funded technical-assistance project: an international consultant, six months, a scorecard in a spreadsheet, and a handover workshop. By the second year nobody has retrained it, and by the third nobody trusts it.
A Wednesday afternoon, no consultant
The head of risk wrote the request in one paragraph:
Predict which active borrowers will miss two consecutive installments in the next six weeks. Use loan history, bKash repayment timing, savings deposits, group repayment records and field visit logs. Test it on periods the model hasn’t seen, including the cyclone quarter.
The Prediction Engine built a weekly history for all 180,000 active borrowers across three years and defined the target strictly looking forward, with every input measured before the prediction window so nothing from the future could leak in. It set up two separate out-of-time tests - an ordinary quarter and the quarter of the cyclone - and reported them separately. That was the most important design choice in the specification. A model that performs in calm weather and fails in a storm is precisely the wrong model for coastal Bangladesh, and an average of the two would have concealed exactly that.
It trained a constrained logistic model and a gradient-boosted model and chose the boosted model on held-out performance. For every flagged borrower it produced the three leading reasons in Bangla for the field officer: last three installments paid in cash on meeting day, previously by bKash two days early; no savings deposit for four weeks; two group members currently behind.
In the ordinary quarter, borrowers in the top risk quintile went on to miss two installments at 5.1 times the rate of the bottom quintile, and the model typically flagged them about five weeks before the Friday list would have. In the cyclone quarter, the ranking weakened markedly. A district-wide shock hits borrowers whose personal behaviour gave no warning, and the engine reported that plainly rather than blending it into a single reassuring figure.
Composite scenario; timings illustrative. Scores trigger a field visit and a conversation only, never a penalty or a refusal.
That weakness became policy rather than embarrassment. In ordinary weeks, the model sets the field priority. After a storm warning, Upokul switches to the geographic exposure query it built in the Analytics Engine the night of signal seven, because in a covariate shock location predicts better than behaviour. Knowing where a model fails is part of what makes it safe to use.
- Calm weathertop quintile 5.1× the bottom
- Covariate shockranking weakens, and the card says so
Composite scenario; figures illustrative. After a signal warning, exposure is managed with the Analytics Engine's geographic query.
What a field officer does with a probability
The managing director set the rule before the first score reached a tablet: a risk score may trigger a visit and a conversation, and nothing else. It cannot trigger a penalty, cannot refuse a repeat loan, and cannot change a borrower’s terms. Those decisions stay with branch managers and follow the existing credit policy.
Each officer now receives up to fifteen flagged borrowers on Monday, with reasons. The early conversations turned out to be different in kind from the Friday-list visits. The borrower is not yet behind, so there is nothing to be embarrassed about. Often the problem is timing: installment day falls before the weekly market where the household sells its produce, and moving it by two days solves it. Sometimes it is a small input loan needed ahead of planting. Over the first two quarters, the share of flagged borrowers who went on to miss two installments fell well below what the model had predicted for them - which is the intended result for a warning that leads to action.
The same engine produced the other forecast Upokul needed after the cyclone: expected collection efficiency by union over the following ten weeks, which now sits beside the union-level recovery alerts and informs provisioning. When the apex lender’s due-diligence team asked how Upokul identifies deterioration early, the answer was the model card itself - cyclone-quarter weakness printed on it - and the lender treated the candour as a strength.
The two head-office analysts who used to compile the monthly pack now maintain the models. Their question this quarter is whether loan purpose should enter the model, or whether it simply stands in for season.
What the model hands each Monday is a list: fifteen names per officer, eleven hundred officers, ninety branches, afternoons already filled with group meetings, and roads that may be under water. Which doors to knock on first, in what order, along which route, is not something a probability answers. It is a planning problem.
Sources
- Alliance for Financial Inclusion (2025), citing MRA data to June 2025: 33.68M microfinance borrowers; MFIs classify loans as non-performing after more than one month overdue.
- Björkegren & Grissen, Behavior Revealed in Mobile Phone Usage Predicts Credit Repayment (World Bank Economic Review; arXiv 1712.05840): highest-risk quintile 2.8× as likely to default; validated out of time.