Barocas and Selbst (2016) catalogue the ways in which a model built from data can disadvantage a protected group without any intention to do so: the outcome may be defined in a way that favours one group, the training sample may under-represent another, the labels may carry past prejudice, and innocuous features may stand in for the protected attribute. All four arise with mobile data. Women in many markets own fewer handsets and transact through a relative's account. Rural users have fewer agents nearby and so fewer cash-in events. A feature such as the share of transactions made late at night may separate occupations, and through them sex or ethnicity, more than it separates good payers from bad.

Evidence from mortgage lending shows that these concerns are real. Fuster et al. (2022) found that moving from a logistic model to a more flexible machine learning model changed predicted default risk unevenly across racial groups, and Bartlett et al. (2022) found that algorithmic lenders still charged minority borrowers more, though by a smaller margin than face-to-face lenders. Removing the protected attribute from the inputs does not solve the problem, since flexible models reconstruct it from correlated features.

Three families of criteria dominate the literature. Demographic parity asks that approval rates be equal across groups. Equality of opportunity asks that applicants who would repay have the same chance of approval whatever their group, and its stricter form, equalised odds, adds the same requirement for those who would default (Hardt et al., 2016). Calibration within groups asks that a given score mean the same probability of default for every group.

These criteria cannot all hold at once. Kleinberg et al. (2017) and Chouldechova (2017) proved that when default rates differ between groups, a score that is calibrated within each group must produce unequal error rates, except in degenerate cases. A lender therefore has to select a criterion, and the selection is a judgement about which kind of error matters most. Kozodoi et al. (2022) examined the criteria in the setting of credit scoring and argued that a single well-chosen criterion can usually be approached at a moderate cost in profit.

The developer first states who is to be scored, for which product, and what counts as default. A definition based on days past due should be checked for groups whose income is seasonal, since a farmer paid at harvest may be late by the calendar and reliable over the season. Consent must be specific to credit assessment, and the data taken should be limited to transaction metadata: amounts, times, types and counterparties in hashed form. Message content and contact lists are excluded. Protected attributes are needed to test fairness even though they are not used for scoring. They should be collected with a separate explanation, stored apart from the modelling data and made available only to those who run the audit.

Features should describe financial behaviour that has a plausible connection with repayment: the regularity of inflows, the ratio of outflows to inflows, the lowest balance in a month, the frequency of bill payment, the length of account history. Each candidate is then tested for its association with the protected attributes. A feature that predicts group membership strongly and adds little to the prediction of default once other features are present is removed. A feature that does both is retained only with a recorded business justification. Features that describe the applicant's contacts, and not the applicant, are best left out, since they score people by the company they keep.

Two models are fitted. The first is a logistic scorecard, whose coefficients can be read and defended. The second is a gradient-boosted tree model of the kind implemented in XGBoost (Chen and Guestrin, 2016). Benchmarking studies find that such ensembles outperform logistic regression in credit scoring, though the margin varies (Lessmann et al., 2015). The flexible model is adopted only if its advantage on held-out data is large enough to matter commercially, because flexibility is also what allows a model to rebuild protected attributes from proxies. Outcomes are observed only for applicants who were granted credit in the past, and those applicants were selected by an earlier rule. Reject inference techniques try to correct for this, with limited success (Crook and Banasik, 2004). Where the lender can afford it, the sounder remedy is to approve a small random sample below the cut-off and observe the result.

Because the criteria conflict, the choice among them is made explicitly and recorded before the audit results are seen. For a product intended to widen access, equality of opportunity has a clear rationale: it asks that creditworthy applicants in every group be equally likely to be approved. Calibration within groups should be checked in every case, since a score that overstates risk for one group overcharges it. The audit then computes, for each group and for intersections where numbers allow, the approval rate, the true and false positive rates and a calibration curve, each with a confidence interval. Small groups produce wide intervals, and the audit should say so and avoid reporting a difference as absent merely because it is not statistically detectable.

If the audit shows a disparity beyond a tolerance set in advance, the developer intervenes at one of the three points shown in Figure 1. Before training, the sample can be reweighted so that group and outcome are independent (Kamiran and Calders, 2012). During training, a constraint can be imposed through the reductions method of Agarwal et al. (2018). After training, cut-offs can be adjusted by group to equalise the chosen error rate (Hardt et al., 2016). The last is the most direct and also the most legally sensitive, because it treats applicants differently by group on the face of the rule, and in some jurisdictions that is prohibited. The procedure therefore prefers the earliest and mildest intervention that brings the disparity within tolerance, and repeats the audit after each change.

An applicant who is refused is entitled in many jurisdictions to the principal reasons, and reasons are also what allow an error to be challenged. For the scorecard they follow from the points lost on each characteristic. For the boosted model, additive attributions such as SHAP values provide a comparable breakdown (Lundberg and Lee, 2017), which should be translated into statements the applicant can act upon. After release the lender tracks approvals, arrears and calibration by group at regular intervals. Mobile money behaviour shifts with tariffs, agent coverage and seasons, so a model that was fair at launch may drift.