Credit intelligence infrastructure

Credit intelligence built to be examined

An explainable machine-learning pipeline designed for risk assessment, validation and deployment across institutional lending environments.

AFRISCORE / TECHNOLOGY

PROCESS

From alternative data to actionable credit intelligence

  1. 01

    Data acquisition

    Authorized financial and behavioural signals are collected through structured integrations. Data availability depends on the borrower, source, market and permissions.

    STEP 01
  2. 02

    Risk estimation

    The scoring engine evaluates available features through an ensemble model and produces a credit-risk assessment. Probability estimates require appropriate calibration and empirical validation.

    STEP 02
  3. 03

    Explainable outputs

    Feature-level explanations help identify the signals contributing to an assessment. Where supported, SHAP analysis and counterfactual methods provide additional insight into model behaviour.

    STEP 03
  4. 04

    Institutional decision

    The lender applies its own underwriting policy to determine approval, loan limits, pricing and terms. AfriScore informs the assessment; the lender retains the decision and credit risk.

    STEP 04

MODEL ARCHITECTURE

Under the hood

AFRISCORE · ENGINE · SPECCONFIGURATION
Model
Random Forest ensemble: 200 trees, maximum depth of 8
Calibration
Isotonic regression using the Pool Adjacent Violators algorithm
Explainability
Kernel SHAP contributions and counterfactual analysis
Data inputs
Alternative financial signals including mobile money, utility payments, airtime and device metadata
Training and validation
Synthetic-data development and stress testing, followed by empirical calibration and validation using real-world data as it becomes available
Deployment
Hosted API or containerized deployment using Docker, subject to integration requirements

SIGNALS & DATA

Signals beyond traditional credit files

The scoring architecture organizes up to 13 structured features across four data families. These signals can help assess applicants with limited or no formal credit history. Availability, relevance and predictive value must be established for each deployment.

5 fields

Mobile money

Transaction activity, transaction volumes, account history, balance behaviour and cash-flow patterns, subject to available data.

SIGNAL CATEGORY

3 fields

Airtime

Recharge frequency, regularity and transaction amounts as potential indicators of behavioural consistency.

SIGNAL CATEGORY

3 fields

Device signals

Selected device characteristics and stability signals, where legally permitted and available. These are supplementary signals, not standalone measures of creditworthiness.

SIGNAL CATEGORY

2 fields

Prior lending

Available information on previous microfinance borrowing and repayment behaviour.

SIGNAL CATEGORY

DECISION & ANALYSIS

Credit intelligence for the lender

  • A numerical credit score
  • A risk classification
  • An estimated probability of default, subject to calibration
  • Feature-level explanations identifying factors contributing to the assessment, where supported by the explainability endpoint

Scoring and explanation requests are separate operations. Response times depend on the endpoint, workload, infrastructure and complexity of the explanation. Risk estimates should not be interpreted as calibrated probabilities until validated against relevant repayment outcomes.

INFRASTRUCTURE

Built for institutional deployment

Hosted API

Connect through a hosted API to evaluate credit intelligence within existing lending workflows, subject to integration, access and deployment requirements.

Private infrastructure

Containerized deployment options can support institutions requiring greater control over data processing, infrastructure, network access and data residency. Actual deployment requirements are agreed with each partner.

VALIDATION

Evaluated before deployment

Feature ablation

Measures how model performance changes when selected input variables are removed, helping identify dependency on specific signals.

Data sparsity

Examines model behaviour when input fields are unavailable or incomplete.

Distribution shift

Tests sensitivity to changes in data distributions and population characteristics across environments.

Explanation stability

Evaluates whether feature-level explanations remain reasonably consistent across resampled datasets and test conditions.

SAFEGUARDS

Governance built into the architecture

Data sovereignty

Deployment options can accommodate institutional infrastructure and data-residency requirements. Data handling depends on the selected architecture and agreement.

Regulatory adaptability

Designed with regional regulatory requirements in mind, beginning with CEMAC, while allowing adaptation to other jurisdictions.

Explainability

Scoring assessments include contributing factors to support review, documentation and responsible use by lending institutions.

Institutional control

AfriScore provides assessments and risk intelligence. Lending institutions retain responsibility for credit policies, decisions and borrower treatment.

Explore the scoring engine

Review the API documentation and submit a test request to examine the assessment workflow.

Open the API reference