Credit Garden update: Charlie Bucket and financial well-being

Credit Garden’s work starts with a borrower whose financial life is more complicated than their credit file. That may be someone with irregular income, limited borrowing history or a business whose cash flow needs a closer look. Our research asks what evidence can help a lender understand that person while giving the borrower a clearer path forward.

We are opening a regular startup update and research series around financial health, financial well-being and the Charlie Bucket AI model for unbanked and underbanked customers. Our work with local fintechs focuses on credit assessment, loan losses and operational efficiency.

What the earlier work established

Credit Garden’s earlier financial-inclusion presentations set out a combination of credit modelling, financial education and guidance. They describe Charlie Bucket as a model entering live testing at the time of that deck. The historical record is useful because it shows the intended problem and the development stage. It does not establish the model’s current default rate or prove a reduction in loan losses.

We will revisit the model’s development and the assumptions behind it, then publish the evaluation with its results. We will also distinguish a synthetic population used for experimentation from records of actual borrowers. A large simulated dataset can help test a process, but its size is not evidence of accuracy on real customers.

A credit decision and a household outcome

A lender wants to know whether a loan is likely to be repaid. A borrower also needs to know whether repayment fits alongside rent, food, transport and unexpected expenses. Our financial-health work keeps those questions together without assuming they have the same answer.

The US Consumer Financial Protection Bureau’s financial well-being questionnaire asks about control over money, resilience to an unexpected expense and the freedom to make choices. Its comparison data concern US adults. We can learn from its measurement approach, but a Caribbean application needs testing for language, relevance and interpretation before treating its benchmarks as local norms.

For Credit Garden, the outcome to measure is whether a service helps people manage their financial lives. Approval volumes and processing time belong in the dashboard. So do repayment stress, complaints, repeat borrowing and the ability to understand and challenge a decision. A faster approval does not demonstrate improved financial well-being.

What the research says about alternative data

Alternative data can add information where a conventional credit file is thin. It can also create new ways to exclude people. The BIS study of Funding Circle and LendingClub examined US small-business loans originated during 2016–2019. The platforms’ internal scores predicted later loan performance more accurately than the traditional scores studied, particularly in areas with higher unemployment. Those findings concern the named platforms and their borrowers.

They provide a reason to test locally, rather than a performance figure Credit Garden can claim. Our first credit research article examines a separate digital-footprint study and explains the difference between ranking borrowers by risk, predicting an accurate probability of default and improving a lending decision.

The role of an AI assistant

Our agent work connects research and day-to-day support. An assistant can help a user organise a budget question, understand the inputs to an assessment and find the source behind an explanation. For a fintech team, it can help locate policy documents and summarise a case for review. These assistant functions are under development.

We apply Adrian Dunkley’s “People First, Problem Second, AI Last” approach here: define the borrower’s need and the decision before choosing the tool. The assistant should make the process easier to understand. It should not invent income, conceal the reason for a recommendation or change a lending policy through a conversation.

What we will measure next

Our research programme will compare scorecards, statistical models and machine-learning models under the same evaluation conditions. The questions include performance on thin-file customers, the stability of explanations, differences between groups and the total effort required to handle a case.

1. Map one lending journey from application to repayment, including the work done by the customer and the reviewer.
2. Record which information was available at the decision date. Keep later repayment information out of the model’s inputs.
3. Audit twenty historical case explanations within two weeks, using authorised, de-identified records. Count unsupported statements and corrections. This suggested process check is separate from a statistical validation of credit risk.

The risks belong in the result

Device details, location and online behaviour may predict an outcome while also reflecting income, unequal access or other sensitive characteristics. Removing an explicit protected field does not remove every proxy. Customers can also be harmed by stale data or by an explanation they cannot contest. Our evaluation therefore needs data minimisation, an appropriate legal basis, checks of group outcomes and a practical route to human review. We will not describe an untested model as fair, compliant or loss-reducing because it uses AI. Any reported commercial benefit must identify the comparison, period, costs and uncertainty.

We will bring these financial-inclusion questions into the portfolio’s support for Caribbean AI Association think tanks and responsible AI discussions. The purpose is to make the evidence behind a financial decision easier for borrowers and lenders to inspect.