What Implementing Alternative Borrower Assessment Actually Taught Us
Author: Priya Maharaj
PEMiG Monday Intelligence Brief | Special Edition Credit Intelligence | Causal AI | African Lending | Financial Inclusion
In the seven months between September 2018 and April 2019, sixty-five digital credit applications exited the Kenyan market. Forty-seven entered. The net effect was not growth — it was churn, an industry cycling through models that could not hold.
That statistic from FSD Kenya’s market monitoring is worth sitting with before discussing anything about alternative borrower assessment. Because what it describes is not a story of financial inclusion expanding and contracting. It is a story of models that worked until they did not, followed by institutions that could not explain why, followed by exits.
The institutions still in the room seven years later and those entering it now with regulated, bank-grade ambitions; did not survive by finding better data. They survived by building better infrastructure around the data they had. That distinction is where this piece starts.
The Velocity Problem Nobody Talks About
Kenya leads Africa in digital lending. The World Bank’s 2025 Global Findex report shows that 32% of Kenyan adults borrow from mobile-money providers, with 25% borrowing exclusively this way. Behind that number is an enormous volume of alternative data; M-Pesa transactions, merchant wallet activity, mobile usage patterns, chama contribution records, all moving through models making credit decisions in real time.
The problem is not the data. It is the velocity at which it changes.
Traditional credit variables like loan repayment history or utilisation ratios evolve slowly. They shift over quarters. Alternative data features move at a completely different rate. Research examining digital credit default in Kenya using the 2024 FinAccess survey found that economic shocks are a primary driver of borrower default; which means that a model trained on mobile money behaviour during a stable period is looking at a fundamentally different distribution than what it will encounter after a fuel price spike or a regional liquidity shock.
The consequence is concept drift: the gradual decoupling of a model’s learned patterns from the current reality of the borrowers it is scoring. In a bureau-based model, this drift is slow and detectable. In an alternative data model, it can happen within a single month, invisibly, long before a monthly portfolio review catches the decline.
A February 2026 ethnographic study following data science and product teams building credit scoring systems for the Kenyan market conducted over nine months between February and November 2025 found that institutions were grappling with the gap between how models behaved in development and how they performed in production, particularly when local economic disruptions caused borrower behaviour to shift faster than retraining cycles could accommodate.
This is the first operational lesson of alternative data at scale: the monitoring infrastructure matters as much as the model. Without real-time feature importance tracking, you will not know the model has degraded until the NPL report tells you three months after the portfolio started moving in the wrong direction.
The Correlation Trap And What It Does to Portfolios
TransUnion Kenya partnered with FICO in February 2025 to launch an advanced credit score incorporating 145 data points and 24 months of behavioural data from telecoms and utilities.145 data points is an impressive number. It is also, depending on how those points are used, a significant source of structural risk.
Complex predictive models excel at finding mathematical patterns that are highly predictive in a backtest. They are considerably less good at distinguishing between patterns that mean something and patterns that happen to correlate with an outcome in the historical data.
The 2024 FinAccess study on determinants of digital credit default in Kenya identified a correlation between betting behaviour and digital credit distress — a finding that immediately raises the causal question: does betting cause default, or do financial stress and betting both arise from the same underlying income shock? A correlative model cannot answer that question. It encodes the correlation and uses it. An institution relying on that model is now declining borrowers based on gambling app usage without knowing whether the signal reflects genuine credit risk or a proxy for something the model cannot actually see.
This is not a hypothetical. It is precisely the kind of hidden bias that gets discovered after a portfolio underperforms and a postmortem team spends three months trying to understand why a seemingly sound model failed a specific borrower segment. By that point, the damage is already in the provisions.
The same logic applies at greater scale when social network data is used for underwriting. The insight that a borrower’s guarantor network carries information about their creditworthiness is correct. The risk is in how that information is processed. If a central borrower in a community network defaults, a correlative model will often downgrade every connected individual — creating a cascade of incorrect declines that locks out creditworthy borrowers based purely on network proximity, not on any assessment of their actual repayment capacity.
The ethnographic fieldwork in Kenya found that data science teams were aware of these proxy and association problems but lacked the interpretability tooling to diagnose them at the individual borrower level; meaning that systematic biases could persist across thousands of decisions before they surfaced in aggregate portfolio data.
What M-Shwari’s October 2025 Outage Actually Illustrated
In late October 2025, M-Shwari; the Safaricom-NCBA Bank partnership that has served millions of Kenyans since 2012 — experienced a significant technical outage. Users across Kenya found themselves locked out of funds, with transactions failing and balances inaccessible for hours.
Get PEMiG’s stories in your inbox
Join Medium for free to get updates from this writer.
The response from users was immediate and instructive. People began splitting their funds across platforms, exploring alternate channels, and keeping small cash buffers; adaptive behaviour driven by the erosion of trust in a single digital touchpoint.
What that behavioural shift looks like in alternative credit data, in the weeks that follow a disruption like this, is a meaningful change in M-Pesa transaction patterns for a large population of borrowers — patterns that have nothing to do with their creditworthiness and everything to do with a platform reliability event they had no control over.
A correlative model sees the pattern change. It does not know why it happened. It updates its risk assessments accordingly. Borrowers who were scored as low-risk before the outage may be scored differently in the following month without any change in their actual financial situation.
This is the operational fragility that black-box alternative data systems create. When external events disrupt the data rather than the borrower, a model that cannot explain its own feature weights cannot distinguish between the two. The portfolio bears the cost of that uncertainty.
The Explainability Requirement Is Not Regulatory Decoration
Experts project that Kenya will move toward a hybrid credit model by 2027, combining CRB data with API-based real-time feeds. The CBK has indicated through consultative papers that it may soon mandate live integrations between CRBs and lenders, a move similar to Rwanda’s 2024 rollout of real-time credit reporting.
That regulatory direction carries an operational implication that is often missed: real-time data integration without real-time explainability creates real-time liability.
When a credit committee can see the specific causal drivers behind a decision — not just the score, but the mechanisms producing it — they can do three things that a black-box system cannot support. They can decouple temporary behavioural shifts from structural default indicators. They can stress-test specific borrower segments against macroeconomic scenarios before the scenarios materialise in the portfolio. And they can explain to a regulator, an auditor, or a partner institution precisely why a borrower was approved or declined — in terms that reflect actual financial reasoning, not mathematical pattern-matching.
The AFI 2025 report on alternative data for credit scoring confirms that digital lenders in emerging markets are adopting these data sources to responsibly expand credit access while maintaining sound portfolio performance; but the “responsibly” in that sentence carries weight. Responsible alternative data underwriting is not just about what data is used. It is about whether the institution can account for what the data is doing inside the model.
The lenders who built sustainable NPL ratios through Kenya’s 2022–2025 rate and currency cycle were not the ones running the most sophisticated predictive models. They were the ones whose credit teams could see and explain what was driving decisions at the individual level and could therefore make precise adjustments before risk became realised loss.
The Structural Lesson
PEMiG’s causal credit scoring engine was built from the understanding that over 80% of adults in Sub-Saharan Africa are thin-file or no-file borrowers — people with real creditworthiness that existing instruments cannot see. The challenge of serving that market is not primarily a data collection problem. It is an interpretation infrastructure problem.
Every institution building or evaluating an alternative data underwriting system should be asking three questions that have nothing to do with data coverage:
Does this model know the difference between a signal that causes repayment and a pattern that merely correlates with it in historical data?
Can this model tell the difference between a borrower who is financially stressed and a borrower whose M-Pesa behaviour shifted because a platform they depend on went offline for six hours?
When a cluster of defaults appears in a specific network or geography, can this system identify the actual causal mechanism in real time or will it take a three-month postmortem to reconstruct what happened?
If the answer to any of these is no, the institution is running alternative data through an infrastructure that was not designed for it. And in a market that moves at the velocity Kenya’s credit environment now moves, that gap will eventually express itself in the portfolio; quietly, then suddenly, in a column of the NPL report that takes longer to explain than it took to create.
Alternative borrower assessment is the future of financial inclusion in East Africa. Its success depends entirely on whether the infrastructure supporting it can account for what it is doing; at the individual borrower level, in real time, across the full loan tenure.
What operational challenges has your institution encountered when moving from traditional to alternative credit data? We want to hear from the practitioners.
📧 felix@pemig.co.ke · 🌍 pemig.co.ke · 📞 +254 745 694 105
PEMiG — Credit Intelligence for African Lenders | Philex Labs Limited | Nairobi, Kenya As featured in Disrupt Africa | African CEO | Fintech Association of Kenya | We Are Tech Africa Selected #3, YASR30 2026 — Yale Africa Startup Review
#PEMiG #CreditIntelligence #CausalAI #AlternativeData #CreditRisk #RiskGovernance #ExplainableAI #AfricanFintech #KenyaBanking #NPLReduction #FinancialInclusion #MSME #EastAfrica #YASR30 #AfricanLenders #ModelRisk #FintechInfrastructure #AfricanStartups
