Context
In a retail bank or a microfinance institution, the cost of risk is the primary driver of profitability. It is the charge the institution records each year to cover expected or realised losses on its loan book, and a single-point increase can absorb the entire margin of a whole portfolio. Data published in 2024 by the UMOA Banking Commission make this point clearly. UEMOA banks report a gross non-performing loan ratio of 8.5%, a provisioning coverage of 61.5% and a stable cost-to-income ratio of 60.8%. The report also notes that the microfinance sector is going through a quality crisis that raises questions about the sustainability of its business model.
Yet, most financial institutions in sub-Saharan Africa still assess credit risk through the individual judgment of loan officers, without a structured quantitative framework. This article explains how to build an industrial-grade credit scoring model, which methodological choices really matter, and what business gains to expect. It is aimed at chief risk officers, data scientists and executives of lending institutions considering this journey.
The daily credit challenge
Every day, a lending institution must decide whether or not to grant credit to thousands of clients. Being too lenient means accumulating defaults. Being too strict means turning away sound clients and losing net banking income. Between these two extremes lies a trade-off to optimise, which comes down to a simple question. What is the probability that this client will default in the coming months?
This is exactly what a PD (probability of default) scoring model estimates. Its business value does not depend on the mathematical sophistication of the algorithm, but on how rigorously the problem is framed, the data is built, the methods are chosen and the outcome is industrialised.
Framing the problem
Before writing a single line of code, three definitions must be locked in.
Default is defined as an arrears of 90 days or more, or an unlikely-to-pay indicator such as a restructuring or a move to legal recovery. This is the Basel definition, usable for IFRS 9 and for pricing. A looser definition at 30 days mainly captures noise and weakens the model, whereas a stricter definition reduces the number of events available for modelling too much.
The performance window is generally set at 12 months, the standard aligned with Basel and IFRS 9. It can however be reduced to 6 months when data depth is limited, or on short-cycle products such as campaign loans, nano loans or microloans with short repayment schedules. This choice determines the historical depth required for modelling.
The scope is finally shaped by a segmentation by client type and possibly by product, since the default drivers of a salaried individual bear no resemblance to those of an SME or of a microfinance client. One model per homogeneous segment remains the rule.
Building the dataset: the real performance lever
Before even discussing feature engineering, serious work on data quality is the foundation for everything else. A model trained on noisy, incomplete or inconsistent data produces unstable scores, regardless of the algorithms used downstream. This requires a robust strategy for data collection, cleaning and control: identifying reliable sources, handling missing values, dealing with duplicates, reconciling client references across systems, detecting outliers, checking temporal consistency. This foundation typically represents several months of work on a first project, but it lastingly conditions every downstream building block of the system.
Once this base is clean, feature engineering becomes the primary performance lever, more important than the choice of algorithm. A concrete example illustrates this well. From the daily movements of a bank account over 24 months, one typically builds three families of variables.
First, on the balance dimension, with rolling averages over 3, 6 and 12 months, the standard deviation, the minimum over 6 months, and the number of days in overdraft. Second, on the flows dimension, with total average monthly credit flows, the regularity of the salary transfer, the debit-to-credit ratio and the 12-month trend. Third, on the incidents dimension, with the number of incidents over 3, 6 and 12 months, the recency of the most recent incident and the cumulative amount.
Each feature encodes a different business signal. Two clients with the same average balance can have completely opposite profiles if one remains stable while the other regularly slips into overdraft. It is the aggregation of complementary signals that produces predictive richness.
The main pitfall to avoid is data leakage. Every feature must be computed strictly from information available before the observation date. Using post-observation data such as recovery actions or restructurings produces a model that shows 0.95 Gini in test and collapses in production.
Choosing the right algorithms
Three algorithms cover the vast majority of use cases, in a complementary rather than competing logic.
Logistic regression
It is the reference production model.
P(Default | X) = 1 / (1 + exp(−(β₀ + Σⱼ βⱼ Xⱼ)))
Each coefficient measures the direct effect of a variable on the log-odds of default, and this interpretability is its main strength. A loan officer can explain a refusal, a regulator can audit the mechanics, and internal audit can decompose every score. It is therefore the algorithm of choice for credit origination and for PDs used in the calculation of IFRS 9 provisions, where transparency is a prerequisite. Its limitation is that it imposes a linear relationship between features and log-odds, which must be compensated by careful feature engineering.
Random Forest
It automatically captures non-linearities and interactions. Its main role is not to be deployed in production, but to serve as a challenger model. Built in parallel to logistic regression, it measures the performance potential of the dataset. When Random Forest gains more than 5 Gini points, it means there are patterns not captured by the linear model, and it becomes necessary to enrich feature engineering. Its second use lies in diagnostics, since the variable importance it produces guides the search for relevant new features.
Gradient Boosting
It currently delivers the best performance on tabular problems. On an early warning model for the early detection of deterioration on the existing portfolio, it can gain 3 to 5 Gini points over logistic regression. On a portfolio of 50,000 clients, this translates into several hundred defaults detected earlier each year, hence several billion CFA francs in provisions avoided through preventive action. It is the algorithm of choice for use cases where performance takes precedence over transparency, such as early warning, fraud detection or behavioural models. In these situations, the model is used to prioritise internal actions rather than to justify a client refusal, and the transparency issue becomes secondary.
In summary, logistic regression is used for origination and IFRS 9 PDs, where transparency and auditability are essential. Random Forest is deployed as a challenger and diagnostic tool. Gradient Boosting is used for early warning, fraud and behavioural models, where performance takes priority.
Validating along three distinct dimensions
Model performance is not just about Gini. Three questions must be answered separately.
Does the model discriminate well? This is measured by Gini and KS. On Individuals, the usual range is between 0.55 and 0.70. On SMEs, we aim for 0.45 to 0.60. On Corporate, a performance of 0.30 to 0.45 is already satisfactory given the naturally limited statistical basis on this segment. These orders of magnitude come from reference literature (Anderson, The Credit Scoring Toolkit; Siddiqi, Intelligent Credit Scoring) and industry practice, and must be adjusted to the African context where data is sometimes less deep.
Are the predicted PDs reliable? This is the calibration question. A model can discriminate well but calibrate poorly, in which case IFRS 9 provisions will be biased and risk-based pricing distorted. This is checked by backtesting class by class, comparing the average predicted PD with the observed default rate.
Does the model remain stable over time? This is measured with the PSI (Population Stability Index), applied both to the global score and to each variable. PSI is the first early-warning system for drift, because it can be measured immediately, without waiting several months to observe defaults. The standard interpretation thresholds are widely accepted in the industry.
Industrialising the model
A model running in a notebook has no value. Five building blocks structure the move to production.
- The scoring engine, exposed as a REST API and deployed as a microservice.
- Full traceability, where every decision is logged with the client identifier, the features, the model version, the score and the timestamp.
- Continuous monitoring, based on a dashboard combining PSI, cohort performance, calibration and override rate.
- Formalised recalibration procedures, with annual intercept adjustment, full recalibration every 2 or 3 years, and a full redesign every 4 or 5 years.
- Governance resting on a Model Committee, strict separation between modelling and validation, a central model inventory and documentation that can be presented to the regulator.
Observed gains
In comparable contexts, retail banks and microfinance institutions in sub-Saharan Africa that have deployed their first industrial-grade scoring system, the observed gains are significant.
The default rate falls by 30 to 40% on scored segments, within 12 to 24 months of go-live. The effect comes from better selection at origination, since clients whose score is below the cut-off are refused or redirected to manual review, and these clients concentrate a disproportionate share of future defaults.
The cost of risk decreases by 1 to 4 points, with a direct impact on net income. For a bank with a loan book of 100 billion CFA francs and an initial cost of risk of 5%, a 2-point improvement represents 2 billion CFA francs of recurring annual gain.
Automation also extends to the entire portfolio of applications, with approval turnaround dropping from several days to just a few minutes. Credit analysts can then focus on grey-area cases and on-site visits, where their added value is greatest.
RAROC improves by 1 to 3 points thanks to risk-based pricing. A client with a 0.5% PD pays a lower rate than a client with a 3% PD, which is both economically fair and competitive on good profiles.
Finally, IFRS 9 compliance is strengthened with reliable and auditable PDs for ECL calculations, without recourse to rough approximations. Decision consistency also improves across branches and officers, eliminating individual biases.
Programme payback is typically observed over 12 to 24 months. The long-term value, made of stronger risk culture, commercial agility and regulatory credibility, is even more structural.
Working with Mansa Research Technologies
Building an industrial-grade scoring system is a demanding undertaking. It combines skills that are rarely found together in a single team, from methodological framing aligned with the Basel and IFRS 9 standards to data engineering on often heterogeneous banking information systems, through statistical and machine learning modelling, software industrialisation, model governance and change management with business teams.
Mansa Research Technologies has the resources and the expertise to carry out this transformation quickly. Our teams bring together data scientists, quant researchers and AI experts trained in applied mathematics, with a dual experience. On one side, that of major European banks, where they designed and industrialised artificial intelligence models within the most demanding prudential frameworks. On the other, that of banks and microfinance institutions in sub-Saharan Africa, whose specificities they know well.
We work across the entire cycle, from methodological framing to building the risk data mart, through segment-level modelling, independent validation, industrialisation of the scoring engine, and the implementation of model monitoring and governance. Our approach favours the transfer of skills to internal teams, so that the system is mastered in-house and remains sustainable and scalable over time.
The ROI materialises from the first year after go-live, with measurable gains on the cost of risk, operational efficiency and the quality of origination decisions.
To get started or explore a specific use case, get in touch.
Sources: UMOA Banking Commission (2024). Annual Report 2024. BCEAO (2024). Annual Report 2023. BCEAO (2025). Report on Banking Conditions in WAEMU – 2024. Anderson, R. (2007). The Credit Scoring Toolkit. Oxford University Press. Siddiqi, N. (2017). Intelligent Credit Scoring (2nd ed.). Wiley. Basel Committee on Banking Supervision (2005). Studies on the Validation of Internal Rating Systems, Working Paper No. 14.