The verdict in three sentences
For a microlender with 60,000 borrowers in Singapore, a custom credit scoring engine costs, as a 2026 order of magnitude, SGD 150,000 to 350,000, depending on the quality of historical data and the depth of integration with your loan management system. The realistic goal is twofold: bring PAR30 below 5% and deliver a decision in under 24 hours instead of 3 to 5 days. Scoring does not replace credit officers or the credit committee: it standardises decisions on small loans and focuses human judgement on borderline cases.
Engine components and their cost
A useful scoring engine combines two layers. The first is a rules engine: eligibility, loan caps, debt service ratio and a mandatory check with the Moneylenders Credit Bureau. The second is a statistical model trained on your own repayment data, which assigns a probability of default to each application.
| Component | Content | Indicative 2026 cost (SGD, ex GST) | Build time |
|---|---|---|---|
| Data migration and cleaning | 3 to 5 years of loans, repayments, arrears | 20,000 to 45,000 | 3 to 5 weeks |
| Rules engine | Eligibility, caps, debt ratio, credit bureau check | 30,000 to 70,000 | 4 to 6 weeks |
| Statistical model | Logistic regression or tree models, calibration, validation | 40,000 to 90,000 | 6 to 8 weeks |
| Officer and borrower app | Online application, document capture, Singpass MyInfo prefill | 30,000 to 70,000 | 6 to 8 weeks |
| Loan system integration | Read balances, write decisions and limits | 15,000 to 45,000 | 3 to 5 weeks |
| Model monitoring | PAR30 by score band, drift, committee reports | 15,000 to 30,000 | 2 to 3 weeks |
Total: SGD 150,000 to 350,000. Annual running costs, model recalibration included, represent 10 to 15% of the build. A lender with only 12 months of clean history should start with rules only, then add the model once 2,000 to 3,000 loans have matured.
On compliance, licensed moneylenders operate under the Moneylenders Act and Rules: interest capped at 4% per month, late interest also capped at 4% per month, a monthly late fee of at most SGD 60 and total charges capped at the principal. The engine must keep the reasoning behind every decision, and lenders that follow the MAS FEAT principles on fairness and transparency gain credibility with partners and funders.
A readable decision grid
A score is only valuable if it translates into simple decisions officers understand. Example of a grid calibrated on unsecured personal and SME loans of SGD 500 to 10,000:
| Score band | Decision | Observed 30-day default | Share of applications |
|---|---|---|---|
| A (750 and above) | Automatic approval, cap raised by 20% | 1.2% | 22% |
| B (650 to 749) | Automatic approval at standard cap | 2.8% | 34% |
| C (550 to 649) | Additional documents, officer decision | 6.5% | 26% |
| D (450 to 549) | Credit committee, guarantor required | 12% | 12% |
| E (below 450) | Reasoned decline | 25% and above | 6% |
With this grid, 56% of applications get an answer the same day. Officers spend their time on bands C and D, where judgement genuinely matters.
What scoring changes on the portfolio
| Indicator | Before | 12-month target |
|---|---|---|
| Average decision time | 4 days | under 24 hours for A and B |
| PAR30 | 7.5% | below 5% |
| Annual credit losses on portfolio | 2.5% | 1.7% |
| Applications handled per officer per month | 120 | 200 |
| Declines with written reasons | under 30% | 100% |
Mini case study
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Wei Ling, chief risk officer of a licensed moneylender in Singapore with 60,000 borrowers, manages a SGD 150 million portfolio. She approves a SGD 250,000 project with SGD 30,000 a year in running costs. To stay prudent, she retains only a 0.3 point drop in losses in year one, instead of the 0.8 point target: SGD 150 million x 0.3% = SGD 450,000. Net gain: 450,000 - 30,000 = SGD 420,000. The build is covered in 250,000 / 420,000 x 12 = about 7 months, before counting the extra borrowers served thanks to 80 more applications per officer.
FAQ
How much history is needed to train a reliable model?
Ideally 2,000 to 3,000 matured loans, including at least 150 to 200 defaults. Below that, start with a rules engine and switch to the model after 9 to 12 months of data collection.
Does scoring work for thin-file borrowers?
Yes, with alternative variables such as employment tenure from MyInfo, CPF contribution history and repayment behaviour on earlier small loans. These borrowers receive a lower starting cap, for example SGD 500, raised after two cycles without arrears.
Can the model discriminate against some borrowers?
The risk exists if sensitive variables are used. The engine excludes race, religion and gender, and a quarterly review compares approval rates across segments to flag gaps above 5 points.
How is borrower data protected?
The PDPA applies, along with the Moneylenders Rules on record keeping. The engine encrypts documents, logs every access and applies role-based permissions.
Who keeps control of the final decision?
The board-approved credit policy sets thresholds. The committee can override a score, but every override is logged and its default rate tracked separately.
Let's scope your project. Send us your portfolio size, available history and current PAR30, and we will price a phased engine, rules first then model, with an indicative budget and timeline. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
