A credit bureau score tells you what happened on credit. A bank statement tells you what is happening right now. A behavioural credit score bridges that gap: it takes three months of raw transaction data and converts it into a structured, explainable risk signal that sits alongside the bureau file, not instead of it.
Below: how the four score families work, what the output looks like, and where this sits in your decision workflow. Whether you assess micro-loans, vehicle finance, or mortgage origination.
The core idea
A behavioural credit score is derived from analysing bank statement transactions, not credit bureau data. It is a form of alternative credit data: using cash conduct rather than repayment history to assess risk. Where a bureau score summarises years of repayment across multiple credit facilities, a behavioural score reads how money comes in, how it goes out, and how the consumer manages the gap between.
The input is three consecutive months of bank statements, uploaded as a PDF or pulled via open banking. AffyScore's extraction engine categorises every transaction: salary deposits, debit orders, loan repayments, gambling transactions, cash withdrawals, returned items, transfers. From that categorised data, the scoring engine calculates metrics across four distinct families and combines them into a single composite score.
The concept is not new. Bank statement scoring has been used in micro-lending for years, driven by Regulation 23A's requirement to verify income and expenses from bank statements. What has changed is the depth of signal extraction and the speed: structured JSON output in seconds rather than manual underwriter review over hours.
The four score families
AffyScore's composite score is built from four weighted families. Each family measures a distinct dimension of financial behaviour. Together, they answer the question every credit provider needs answered: if we extend credit to this person, what is the probability that they will meet their obligation?
Income: 30% of the composite score
The Income family does not just measure how much a person earns. It measures how reliable that income is.
Three metrics drive this family:
- Salary recurrence: Does a consistent salary deposit appear each month, from the same source, within a predictable date window? A salaried employee at a large corporate will score differently from a gig worker receiving irregular deposits from multiple sources.
- Income stability (coefficient of variation): The CoV measures month-to-month variance in total income. In AffyScore's default model, a CoV below 0.15 signals highly stable income, while a CoV above 0.35 means the consumer's income swings by more than a third between months, a material risk factor for any fixed-instalment product.
- Irregular income share: What proportion of total income comes from non-salary sources? A consumer earning R18,000 in salary and R2,000 from ad hoc transfers has a 10% irregular share, which is manageable. A consumer earning R12,000 from salary and R14,000 from sporadic deposits is far more exposed to income disruption, even though their total income is higher.
High income with high variance is a worse signal than moderate income with low variance. A vehicle finance provider needs to know R4,200 will be there on the 1st of every month for 60 months. The Income family tells them whether it will be.
Cash Buffer: 25% of the composite score
The Cash Buffer family measures resilience. Not just whether the consumer can pay today, but whether they can absorb an unexpected cost without defaulting.
- Average daily balance: The mean balance across all days in the three-month window. This is not a snapshot. It smooths out the payday spike and the month-end trough to give a true picture of liquidity.
- Low-balance days: The number of days where the account balance drops below a configurable threshold (AffyScore defaults to R500 or 5% of monthly income, whichever is higher). A consumer who spends 20 out of 30 days near zero has no room for a missed payment or an emergency expense.
- Month-end resilience: What does the account look like on the 25th through 31st? This captures whether the consumer reaches the end of the month with anything left or whether they are running on fumes before the next salary lands.
Cash buffer is a signal bureau data structurally cannot capture. A consumer can have a clean bureau record (no missed payments, no judgements) while running their account to zero every month. The bureau sees only that obligations were met. The bank statement sees how close they came to not meeting them.
Discipline: 25% of the composite score
Discipline examines how the consumer manages existing financial obligations. This is where behavioural scoring and affordability assessment under Reg 23A overlap most directly.
Start with returned debit orders. In SA credit risk practice, dishonoured debits are widely regarded as a leading indicator of near-term default. Each return compounds the consumer's financial pressure with bank penalty fees, and AffyScore counts both frequency and recency. A single return six weeks ago carries a different weight from three returns in the last month.
Then lender stacking. Multiple active short-term lenders appearing on the same statement signal distress borrowing. A consumer servicing obligations to Boodle, Lime24, and FairMoney simultaneously carries a fundamentally different risk profile from a consumer with a single personal loan at a major bank, even if the total debt-service amount is similar.
Finally, the debt-service ratio: the percentage of gross income consumed by identifiable debt repayments. This feeds directly into the Reg 23A affordability calculation. In AffyScore's default configuration, a ratio above 45% signals limited capacity for additional obligations, though the threshold is configurable per credit product.
This is the family where the bank statement most directly contradicts what a self-declared application form claims. Consumers routinely understate existing obligations. The bank statement does not.
Red Flags: 20% of the composite score
Red Flags captures behavioural signals that predict default but remain invisible to traditional assessment. Not disqualifying on their own; these are weighted risk inputs.
Gambling share-of-wallet is the headline metric. AffyScore calculates the percentage of gross income directed to gambling merchants (Betway, Hollywoodbets, Sportingbet, and similar) using both merchant category codes and transaction description matching. Absa's personal banking unit has begun incorporating gambling trends into lending risk assessments, describing the spend pattern as a significant predictor of delinquency. The scoring engine treats gambling share as a sliding scale, not a binary flag.
Cash withdrawal dominance creates a different kind of risk. When a disproportionate share of outflows is ATM withdrawals, the scoring engine loses visibility into actual spending. High cash withdrawal ratios also correlate with informal debt obligations that never appear as structured transactions, with mashonisa (informal) loans being a common SA example.
Transaction anomalies round out the family. A salary deposit that disappears in month three. A new high-frequency debit order from an unrecognised lender. Outflows that spike 40% above the three-month average. These pattern breaks do not prove distress, but they flag accounts that warrant closer review.
Red Flags carries the lowest weighting, but it catches risk that other methods structurally cannot see. A bureau score has no visibility into gambling spend. A payslip check cannot detect cash withdrawal patterns.
Score range and recommendation bands
AffyScore outputs a composite score on a 300–850 range, aligned with TransUnion's FICO Score 6, one of the most widely used credit scoring scales in South Africa. Credit analysts familiar with FICO-based bureau scores can interpret the output without retraining. (Other SA bureaus use different ranges; AffyScore chose the 300–850 scale for its familiarity.)
In AffyScore's default configuration, the score maps to three recommendation bands:
- Clear (typically 680+): The consumer's behavioural profile supports the proposed credit. Income is stable, buffer is adequate, discipline metrics are healthy, no material red flags. The credit provider can proceed with confidence, subject to their own policy rules.
- Caution (typically 500–679): The behavioural profile shows mixed signals. Perhaps income is strong but the gambling share is elevated, or the buffer is thin despite good discipline. The consumer may still qualify, but the credit provider should apply additional scrutiny: a higher deposit, a shorter term, or a manual review.
- Review (typically below 500): Multiple behavioural indicators suggest elevated risk. This is not an auto-decline. AffyScore never declines; it advises. The recommendation is that this application should receive human review before a decision is made.
AffyScore is an engine, not a decision. The output is a structured risk signal with plain-English explanations for each band. The credit provider's underwriter, or their automated rules engine, makes the call.
Reason codes: every point is traceable
Credit providers face a recurring problem with traditional scoring: the consumer gets a number with no explanation, and the credit provider cannot articulate the factors to the NCR if challenged. Under Section 80(1)(a) of the NCA, a credit agreement is reckless if the credit provider failed to conduct the assessment required by Section 81, and an assessment that cannot be reconstructed or explained is difficult to defend. Opacity is a compliance risk, not just a UX problem.
AffyScore addresses this with reason codes. Every score includes a ranked list of the factors that influenced it, expressed in plain English with the underlying metric:
- "Income CoV exceeds 0.35: monthly income varies by more than 35%"
- "Gambling share-of-wallet above 8%: R3,200 of R38,000 gross income directed to gambling merchants"
- "3 returned debit orders in the last 30 days"
- "Lender stacking detected: 4 active micro-lenders identified"
- "Average daily balance below R800 for 2 of 3 months"
This audit trail serves three purposes. The underwriter sees exactly what drove the score. If the consumer disputes the outcome, the credit provider can point to specific, verifiable transaction patterns. And when the NCR audits the assessment process, reason codes demonstrate that the decision was grounded in objective data, not a black-box algorithm.
How it differs from a bureau score
A bureau score and a behavioural score answer different questions. The bureau asks: based on years of credit history, how likely is this person to default? The behavioural score asks: based on three months of cash conduct, do they have the capacity and stability to meet a new obligation?
Both are valuable, and neither is sufficient alone.
The bureau misses anything not yet reported. A consumer who lost their job last month still shows a clean file, until the first missed payment lands 60 days later. A consumer who started gambling heavily in January will not show bureau distress until defaults filter through in March or April. A consumer stacking micro-loans from five lenders may not show all of them if some lenders do not report to all bureaus.
The behavioural score catches these signals in near real-time. The salary that stopped appearing. The gambling transactions that started. The new lender debit orders that multiplied. All visible on the bank statement weeks or months before they manifest as bureau events.
Conversely, the behavioural score does not capture long-term repayment track record. A consumer with a perfect 10-year history who hits a rough patch for two months will score poorly on behaviour but brilliantly on the bureau. The strongest credit decision uses both.
Configurable weightings
The default weighting (Income 30%, Cash Buffer 25%, Discipline 25%, Red Flags 20%) reflects a balanced risk appetite suitable for most unsecured lending. But different credit products carry different risk profiles, and AffyScore allows credit providers to tune the family weights to match.
A payday lender advancing R3,000 for 30 days might weight Red Flags and Discipline higher, because income level matters less than income timing and existing obligation load for a short-term product.
A vehicle financier extending R280,000 over 60 months might weight Income at 35% and Cash Buffer at 30%, because a five-year product demands assurance that earning capacity and financial resilience will persist.
A debt counsellor performing a pre-assessment wants the full obligation picture and collection date alignment for restructuring. They might weight Discipline at 40%, because the question is not "should we lend?" but "how do we restructure?"
The API accepts custom weights per request, so a single integration can serve multiple product lines with different risk configurations.
Where it fits in the credit decision
AffyScore is a supplementary pre-screen. Section 81 of the NCA requires credit providers to assess a consumer's financial means, prospects, and obligations before entering into a credit agreement. AffyScore sits between the application and the final credit decision, providing the structured data for that assessment. The typical integration:
- Consumer applies and uploads a bank statement (PDF or open banking connection).
- AffyScore extracts, categorises, and scores the transactions. The API returns a structured JSON decision pack: composite score, family scores, reason codes, income summary, obligation mapping, and a Recommendation (Clear / Caution / Review).
- The credit provider's system ingests the decision pack alongside their bureau pull, application data, and internal policy rules.
- The underwriter, or the automated rules engine, makes the final decision.
AffyScore does not replace the bureau or the credit provider's own scorecards. It provides a structured, explainable risk signal derived from data the credit provider already collects but historically could not analyse at scale. With over 18 million credit applications per quarter in South Africa, the volume demands automation that manual statement review cannot sustain.
For a detailed look at how bank statement data feeds into the Reg 23A affordability calculation, see our guide on bank statement affordability in practice.
The bottom line
A behavioural credit score reads what the consumer actually does with their money, not what they did with credit years ago. It captures income reliability, financial resilience, obligation management, and high-risk behavioural patterns, all from three months of bank statement data.
Every point on the score has a reason code. It is transparent by design. Weights adjust to match your risk appetite. And it advises, never decides. It adds a layer between a bureau pull and a lending decision: one that surfaces signals the bureau structurally cannot see.
For credit providers who already collect bank statements under Reg 23A, the data is there. The question is whether you are extracting the signal from it.
