01 — Diagnosis: Growth that masks the deterioration of risk
At year-end 2024, WAMU microfinance counts 533 MFIs and 19.1 million clients, for a loan book of XOF 2,695 billion (+5.3% year on year). But the average loan per client is flat: the sector is growing by multiplying files, not by deepening relationships. And more files to assess with the same methods means more poorly controlled risk. That is what the quality indicators reveal.
| INDICATOR | 2023 | 2024 | NORM |
|---|---|---|---|
| Gross portfolio-at-risk rate | 4.4% | 6.6% | ≤ 3% |
| Provisioning rate | 12.4% | ~12% | — |
| Capital adequacy ratio | 14.5% | 13.5% | ≥ 15% |
| Net income (XOF bn) | 23.5 | 16.7 | — |
None of these lines is reassuring. On large MFIs, the gross portfolio-at-risk rate jumped from 4.4% to 6.6% in a year, more than double the 3% norm set by the regulator: the share of loans in difficulty is rising at a pace the sector no longer controls. An impaired loan is not yet a loss, since part of it will be recovered or covered by collateral. But the provisioning rate stays stuck around 12%, which shows that institutions are poorly prepared for the fraction that will tip into default, and on a portfolio deteriorating at this pace, that fraction promises to be substantial.
The capital adequacy ratio, for its part, sinks further below the 15% threshold: already breached in 2023 at 14.5%, it falls to 13.5% in 2024. This threshold is a valuable safeguard, calibrated so that equity can absorb the unexpected losses of a microfinance portfolio. Drifting away from it means watching the margin of safety thin at the worst possible moment, because the share of impaired loans that ends in default will, on the day it is recognised, fall on that same equity. Net income, finally, melts by 29% in a year. Across all MFIs, the portfolio-at-risk rate even reaches 8.9%, against 6.9% a year earlier. This is the whole point of the title: the next wave of losses is not a distant risk, it is already inscribed on the books, waiting to be recognised.
The regulator has already begun to act: 9 MFIs under provisional administration and 5 licence withdrawals in the fourth quarter of 2024 alone, along with reprimands and warnings. But sanctions almost always come after the deterioration, when it is too late. The real question for a leader is not "am I compliant today?" but "can I see deterioration coming before it becomes irreversible?".
02 — Lessons from the failures: Two kinds of failure, two answers
Not all failures are alike, and the remedy differs according to their nature.
UNACOOPEC-CI — Governance & Fraud
Côte d'Ivoire's leading network, up to 81% of the sector's book. 2012 BCEAO audit: embezzlement and falsified entries. XOF 11bn in losses, negative equity of −XOF 20bn.
ASUSU SA (Niger) — Governance & Fraud
The country's largest MFI, "virtually bankrupt" according to the supervisor in 2018. Accounts seized, prudential ratios breached, diversion of funds earmarked for rural credit.
These collapses stem from governance and internal control: they are prevented through clear processes, separation of duties and audit, not through a model. The second family of failures, by contrast, is the most frequent and the least visible.
The most common failure — the silent drift
Sound institutions whose portfolio deteriorates slowly to breaking point, for lack of an early-warning system. This is the case for most of the small structures pulled from the market, and it is exactly the family of failures that a scoring and monitoring framework addresses directly.
03 — Competitive pressure: A window that is closing
The ecosystem is reshaping fast: more than 130 fintechs, 173 million e-money accounts (+25.6% in 2024) and the launch of the instant payment platform (PI-SPI) in September 2025, which opens a pool of transactional data usable to score in seconds. The shift is concrete: Wave, holding an e-money institution licence since 2022, created in October 2025 a commercial bank in Côte d'Ivoire set to grant credit.
The threat is structural. A player able to score finely captures the most creditworthy profiles and leaves the rest to the institution that cannot sort its risks: this is adverse selection. Without an evaluation capability, the MFI carries rising risk for falling returns.
Yet nothing is lost. MFIs hold an asset that new entrants do not: the repayment history of their own clients, accumulated year after year. This is precisely the raw material needed to build solid scoring models suited to their clientele. It still has to be put to work, which means launching a data strategy now, hardening the technical foundation and acquiring rigorous models. These are the conditions for turning a dormant asset into a competitive advantage, and they are exactly what the rest of this note sets out.
04 — What scoring changes: From measuring risk to deciding
A scoring model assigns each file a probability of default (PD), that is, the probability that a borrower stops meeting their instalments. Where craft-based evaluation rests on an analyst's impression, subjective by nature and hard to trace, the model produces a homogeneous, reproducible and explainable measure. It is this objectification of risk that makes a credit decision consistent from one file to the next.
What artificial intelligence brings
The principle of a machine-learning model is simple to state. From the institution's actual repayment history, the algorithm learns the combinations of characteristics that preceded past defaults, then applies that knowledge to each new file to estimate its risk. Where a fixed scorecard applies the same rules to everyone, a model trained on the MFI's own data picks up signals that manual analysis leaves aside, such as the regularity of mobile-money flows, the seasonality of an activity or repayment behaviour over previous cycles.
The available techniques range from the simple scorecard to logistic regression, up to more powerful models such as random forests or boosting. But an algorithm's raw performance is not enough. What makes the difference is the rigour with which the model is built, the reliability of its move into production and the quality of its monitoring over time, because a poorly designed or poorly monitored model produces wrong decisions at scale.
Setting an approval threshold
The decisive contribution is not the score, but the decision it enables: one defines an approval threshold below which a file is not financed. This threshold makes an explicit trade-off between risk and return, and becomes, for the first time, steerable.
| THRESHOLD | FILES FINANCED | EXPECTED DEFAULT |
|---|---|---|
| Permissive (PD ≤ 25%) | 820 | ~11% |
| Balanced (PD ≤ 12%) | 610 | ~6% |
| Prudent (PD ≤ 6%) | 410 | ~3% |
The tighter the threshold, the lower the volume, but expected default falls faster still. Management chooses the balance point knowingly. (Illustrative figures.)
Repeated across the whole portfolio, this discipline "cleanses" it without blunt refusals: the default rate falls without serving fewer clients. At the WAEMU scale, each point of portfolio-at-risk avoided represents on the order of XOF 27 billion in losses spared, to which are added assessment times cut to a few minutes, a per-file cost divided, and better-sized provisions that bring capital back toward its norm. The model does not, however, replace close knowledge of the client: the loan officer's proximity grasps what escapes the figures. Data and field go together.
05 — Implementation: A roadmap, not a standalone tool
A credible framework follows a sequence where each step reinforces the next.
- Scoring — Statistics and expert judgement, segmented by profile.
- PD — Calibrated on internal history and the credit bureau, at the right horizons.
- Limits & thresholds — Derived from the PD and repayment capacity.
- Provisioning — Forward-looking, based on expected loss.
- Model governance — Backtesting, validation, drift monitoring.
- Skills — Teams trained to read and steer the score.
06 — Realities on the ground: A model is only as good as the process that carries it
The transformation plays out in the daily frictions that no algorithm fixes on its own: the paper file that takes weeks to circulate, the multi-borrower client invisible for lack of continuous credit-bureau lookups, spreadsheet monitoring recomputed once a month, commercial incentives set on volume rather than quality. Placing a good model on a faulty process only speeds up the errors.
Two conditions therefore frame success. Process first: digitise the file, integrate the credit bureau in real time, smooth the decision circuit, align incentives with quality. Digitalisation and scoring complement each other, since one provides the vehicle and the other gives it direction. Skills next: a score has value only if teams can read it, explain it and spot its drift. Process and people are not the backdrop of the transformation, they are its foundations.
Sources: BCEAO — Situation of microfinance in the WAMU at 31 December 2024 (April 2025). WAMU Banking Commission — 2024 Annual Report. WAEMU Credit Information Bureau — Cotonou, October 2025. Accion — Credit Scoring, Risk Management Tool Guide No. 3 (January 2024). Wave: e-money institution licence, 14 April 2022 (OSIRIS); Agence Ecofin, October 2025.