AI-Powered Credit Scoring for the Thin File and Informal Economy

One of the most conservative aspects of finance, the decision to credit someone, is being transformed by AI-driven credit scoring. This is a gradual change for individuals whose credit history is rich and who have been with banks over a long period of time. It could be life-changing, however, for thin-file borrowers and those who are in the informal economy. They can finally be detected as opposed to being invisible to the system. This article discusses the application of alternative data to create AI-based credit scores for individuals and small businesses that lack a traditional credit record, the risk of unfairness and bias when bureau data is unavailable, and the regulatory drive to create explainable AI in the underwriting of underbanked populations.

Thin File and Informal Economy Problem

Conventional credit ratings presuppose some sort of financial existence. They assume that an individual has a bank account, formal financial products, and has borrowed a loan or used a credit card previously. They presuppose employers operate payroll in a formal way and merchants work in the visible part of the economy — the documented part. Practically, a colossal proportion of the world is not so. Young adults often have no loans or cards. Migrants can possess good credit backgrounds in their countries of origin and nothing in their new ones. Most of their transactions are done in cash or in digital platforms that do not report to bureaus: gig workers, street vendors, informal shopkeepers, and a great number of micro-entrepreneurs. Where there are bureaus, even their coverage may be superficial or biased toward urban, formally employed populations. The bureau file of such applicants appears blank or almost blank to lenders. As risk teams are trained to trust bureau data, they make mistakes in favor of caution. The result is predictable: increased rejection, narrowed limits, increased prices, or total exclusion.

These borrowers are not necessarily riskier; it is just that the system is deaf and blind to the signals that actually characterize their financial lives. The basic concept of the application of AI to credit scoring here is straightforward. Where bureau statistics are lacking or too sparse, seek elsewhere. There are numerous digital footprints in modern life. When such footprints are gathered in a responsible manner with consent and converted into more organized signals, they can tell a lot about the stability of a person, their earning potential, and their chances of repaying. One of the first and most valuable sources is often telecom data. Mobile operators understand how someone fills prepaid balances on a regular basis, whether they use the same number over years or switch frequently, whether they are steady or haphazard in their activity, and whether they acquire data packs of the same size. An individual who keeps one number over time, reloads the number, and exhibits consistent patterns of utilization is generally more deeply embedded within a community and more consistent in their behavior as compared to someone who drops or swings in utilization. Whether there is stability is associated with reduced credit risk.

Another source of power is e-commerce and data from digital platforms. Little can be contained in the bureau file of a ride-hailing driver, but a platform can access counts of trips, income per week, cancellation data, customer reviews, and duration of the driver. A micro-merchant as a seller in a marketplace leaves behind a history of orders completed, refunds made, complaints brought up, stock-outs, and growth patterns. In the case of informal businesses, platform data can be used as the nearest equivalent to official financial statements. Next, there is a bank account, digital wallet, and open banking API cash-flow data. Although a borrower may lack a long credit history, he or she also tends to have an account where salary, gig income, remittances, or business revenue is deposited. Through the analysis of time-based inflows and outflows, lenders can estimate the common income, its variability, whether it has buffers or not, and what portion of the income has already been allocated to recurring expenses like rent, utilities, and existing debts. In the case of the underbanked borrower, cash-flow underwriting is often more reliable than the traditional scorecard, which relies so much on past loans. One more layer is provided by payroll and employment APIs.

In situations where employers are connected to payroll services, lenders are able to confirm employment, monthly earnings, the duration of employment, and compensation changes. For those with several part-time jobs, this composite image will be much more informative than one pay slip. Lastly, with proper usage, behavioral and device-level data can be used to assist in both fraud and risk estimation. The duration of time an individual has been using the same device, the regularity of their login locations, how they use the app over the months, as well as the time of day they usually do the transactions, may provide indicators of genuineness and steadiness. These signals should be handled with care so as to prevent proxy discrimination, though they can be of useful support. All these sources are connected by the fact that they tell about the real life of a person and how he/she lives, how he/she earns and pays even when it is clear that he/she has never laid his/her finger on a credit card in his/her life.

\

How AI Transforms Sloppy Signals into a Score?

These other sources of data are densely populated and unstructured. The structure of telecom logs, platform events, bank transactions, and device telemetry is not that of a traditional bureau report. They are loud, dimensional, and filled with patterns of idiosyncrasy. At this point, AI, in particular modern machine learning, is necessary. The common lifecycle begins with aggregation of data. Lenders have access to telecom partners, open banking feeds, payroll APIs, and platform partners on the condition of data protection laws and direct consent of the customers.

They absorb raw data into safe environments and normalize it. Phone recharge activities, wallet credit, and e-commerce orders are converted to time series that have regular formats. Unnecessary anomalies and duplicates are eliminated and missing values are processed. Out of this, features are constructed by data scientists. They create summary variables rather than merely feeding all the raw transactions into a model: average monthly net cash flow; the share of months where the savings are positive; the longest consecutive period of no payments to creditors; the months of under-earnings; growth or decay of platform earnings; variability of working hours; permanence of location week by week.

These attributes are trying to squeeze the economic life of an individual into numbers that can be digested by the model. Gradient boosting trees, random forests, and neural networks are then machine learning algorithms that are trained on historical data where the outcome is already known. In the case of credit scoring, the result is usually a default by the borrower over a specified period of time, say six or twelve months. The model gets to know combinations of features that indicate more or less risk. Patterns found among human underwriters would not have been identified by human discernment, like minor interactions between the volatilities of cash-flows and platform tenure. Validation is critical. The model is applied to data that it was not trained on so that its performance is real and not a result of overfitting.

Measures such as AUC, Gini coefficient, and Kolmogorov–Smirnov statistics are used to measure the power of discrimination, whereas calibration plots indicate whether the predicted probabilities are identical to the actual default rates. In addition to headline figures, lenders need to look at performance based on segment: new-to-credit versus experienced borrowers, various occupations, regions, and income bands. Upon deployment, the model will then rate new applicants on the fly, and a response will be provided within a few seconds. The process cannot end there. Statistics change with time, platforms evolve their policies, and macroeconomics evolve.

\

:::tip This story was distributed as a release by Sanya Kapoor under HackerNoon’s Business Blogging Program.

:::

\