Situation
CreditBook wanted to improve the quality of lending decisions using the information already available across its credit and financial-data platforms.
Existing models provided a starting point, but the available data contained considerably more potential predictive information than was being used in the lending decision. The challenge was to determine which signals were meaningful, translate them into defensible risk indicators, and make the resulting model practical for operational use.
GTP’s work
- Assessed the existing decision models, available data, and data flows, then designed the pipelines required to make relevant information consistently available for modelling.
- Performed deep analysis of the underlying data and derived candidate indicators associated with borrower behaviour and lending outcomes, identifying combinations of signals capable of distinguishing stronger and weaker credit risks at different risk tolerances.
- Developed and calibrated an explainable decision-tree and scoring model, tested it against more than 100 client cases, and built the API layer required to make scoring available to operational applications.
Result
The new proprietary model substantially outperformed the existing models in testing and gave CreditBook a stronger quantitative basis for distinguishing lending risk.
The resulting architecture also moved the model beyond analysis: scoring can be invoked through an API, allowing the decision logic to become part of the operational lending process and to be recalibrated as additional evidence becomes available.
Related capabilities