Picture a bank credit officer facing a straightforward problem: the applicant sitting across the desk has a steady salary, pays utility bills on time, and has never missed a rent payment — but has never taken out a loan. The traditional scoring model sees almost nothing. No credit history means no score; no score means no credit. This circular logic keeps millions of otherwise creditworthy individuals at the margins of the formal financial system, and it is a challenge that Turkish banks are beginning to take seriously as alternative data sources become more accessible and analytical tools more capable.
Machine learning, in this context, refers to a family of statistical methods that identify patterns in large datasets and use those patterns to generate predictions for new observations. In credit scoring, the approach moves beyond repayment history alone and incorporates behavioral signals: the regularity of account inflows, the timing of bill payments, the stability of monthly cash flows. Classical methods such as logistic regression and decision trees have been part of credit risk practice for decades; machine learning extends these tools to handle more variables, more complex interactions, and larger training samples. The concept is not new to finance globally, but its practical application in Turkish retail lending is gaining momentum through 2013 and into 2014.
The core logic of alternative data scoring rests on a reasonable premise: how a person manages their financial obligations — even small, recurring ones — is correlated with how they will manage a loan. A model trained on thousands of accounts can learn which combinations of behavioral variables predict repayment reliability, and then apply that learned structure to applicants who would otherwise receive no score at all. This opens a meaningful opportunity for young professionals, self-employed individuals, and salaried workers who simply have not yet entered the credit system. From a portfolio growth perspective, this is an underserved segment with real demand.
The operational case is equally concrete. Traditional credit assessment involves analyst time, document collection, and manual review — a process that does not scale efficiently at high application volumes. A model-based pre-screening layer can sort applications into automatic approval, manual review, and early rejection categories with far less human intervention. For a bank managing tens of thousands of monthly consumer credit applications, the reduction in processing cost per decision is measurable. A manager running a straightforward ROI calculation will find that even a modest improvement in throughput, combined with a reduction in default rates among newly scored segments, justifies the model development investment within a reasonable time horizon.
The fairness dimension, however, cannot be treated as secondary to technical performance. A model trained on historical data inherits whatever biases are embedded in that data. If past lending decisions systematically disadvantaged certain geographic regions, income brackets, or demographic groups, a model trained on those outcomes will reproduce — and potentially amplify — the same patterns. A customer in a district with historically higher default rates may find their application penalized not because of their own behavior, but because of where they live. This is not a hypothetical edge case; it is a structural risk in any scoring system that relies on population-level patterns to make individual decisions.
Turkey’s regulatory environment adds a practical constraint that shapes how these models can be deployed. The Banking Regulation and Supervision Agency’s expectations around credit decision processes require that decisions be explainable and, when challenged, justifiable to the applicant. Telling a customer that the model decided is neither legally sufficient nor commercially sustainable. This pushes lenders toward model architectures where the contribution of individual variables can be traced and communicated — interpretable models rather than opaque ones, at least for customer-facing decisions. The technical capability to build a complex model and the institutional readiness to operate it responsibly are two different things, and the gap between them is where most implementation difficulties arise.
For a bank executive evaluating whether to invest in machine learning-based scoring, the defining question should be framed clearly before the project begins: is the goal to find creditworthy customers the current system cannot see, or to process existing applicants more efficiently? Both are legitimate objectives, but they require different data strategies, different model designs, and different success metrics. A project that conflates the two will likely underdeliver on both. The prerequisite conditions — a well-maintained transaction data infrastructure, a team capable of building and validating statistical models, and a compliance framework that accounts for explainability requirements — need to be assessed honestly. Where those conditions are not yet in place, the more productive near-term investment is building the data foundation rather than rushing to the model.
This article was originally written in Turkish by Gökhan MERCANOĞLU on May 5, 2014 and has been automatically translated into English and other languages using machine translation.