A credit scoring model for microlenders: where to start

Many microlenders start out with lending decisions made by a loan officer’s judgement plus a couple of hard rules. While the portfolio is small, that works. As volumes grow, decisions become inconsistent and portfolio quality depends on who happens to be on shift. A scoring model makes decisions repeatable and measurable. Here is a practical sequence.

The data you need

A model learns from history: who you lent to and how they paid. The minimum set looks like this.

• Application data as of the application date: age, employment, income, region, marital status, length of employment.
• Loan parameters: amount, term, product, acquisition channel.
• Repayment history: schedules, actual payments, days past due for each loan.
• External data where it is available and lawful: credit bureau, the client’s history with you, transaction data.
• Rejected applications — you need them to understand who you are not seeing in the portfolio.

Data must be captured as of the decision date and not overwritten later. If income was updated in the system six months after disbursement, the model will learn from information that did not exist when the loan was approved.

Define what a “bad” client is

Before building anything, agree on a default definition — for example, more than a set number of days past due within a set period after disbursement. The definition must fit the product: it will differ for short-term loans and one-year loans. This is a decision for the business and risk management, not just the analyst.

A simple scorecard or machine learning

For a first model we usually recommend a classic logistic-regression scorecard. The reasons are practical:

• the credit committee understands it, since each characteristic adds a clear number of points;
• it is easy to explain to a client, an auditor or a regulator;
• it is stable on small samples;
• it is simple to plug into any decision engine.

Machine learning models such as gradient boosting are often more accurate, especially with plenty of transaction or behavioural data. But they are harder to explain and monitor. A sensible path is to start with a scorecard, then test whether ML gives a meaningful lift on your data. If the lift is small, simplicity wins.

Validation

The model has to be tested on data it did not see in training: a hold-out sample and, ideally, a later time period. Several things are checked.

• Discriminatory power: how well the model separates good clients from bad (Gini, KS).
• Calibration: whether predicted default probability matches actual default rates by group.
• Stability: whether score and characteristic distributions shift between periods.
• Common sense: whether each characteristic moves the score in a logical direction.

Validation is best done by people who did not build the model. For a small microlender, that may be an external specialist.

Cut-offs

The model produces a score, but the business needs a decision. Three zones are typical.

• Approve — automatically or through a simplified procedure.
• Refer — grey-zone applications where a specialist makes the call.
• Decline — risk above the acceptable level.

Cut-offs are set from product economics: how much a good client earns, how much a bad one loses, what a manual review costs. Move the cut-off and you can see how approval rates and expected losses change. That decision belongs in a conversation with the commercial side, not just within risk.

Embedding it in the lending process

A model only helps once it is built into the process. Decide where the score will be calculated: in the loan origination system, a separate decision engine or the core system. Test the calculation on historical applications and confirm the score in the system matches the score from the model. Train loan officers: they need to understand what the score means, when a manual override is allowed and how to justify it. At first, it is useful to run the model in “shadow” mode and compare its decisions with current ones.

A good scoring model is not the most accurate one. It is the one people understand, apply correctly and check regularly.

Monitoring and drift

Once live, a model starts to age. The economy, acquisition channels, products and client behaviour all change. Set up regular monitoring:

• monthly — score distribution and the population stability index (PSI);
• as loans mature — actual default rates by score band;
• the share of manual overrides and how those loans performed.

Agree in advance which values will trigger recalibration or a rebuild. Then it becomes planned work rather than a reaction to rising arrears.