Upin registers patent converting generative AI confidence into actual default rates

A technology that separates the structural gap between the linguistic confidence of generative artificial intelligence (AI) and the probability that a borrower will actually fail to repay debt, and converts it into measured default rates, has been registered as a patent. It adopts a dual structure that retrains only a lightweight risk calibration function without retraining the large language model (LLM) itself each time, in an attempt to simultaneously control the cost of quantifying credit risk and the error in delinquency prediction.

Upin, a financial AI company led by CEO Lee Chul-won, said on Nov. 10 that it has completed patent registration for a method of calculating and calibrating loan review delinquency probability using log probabilities of loan review response tokens from generative AI, and a computing device for executing the method.

Existing credit scoring models have calculated borrower default risk mainly from structured data such as income and financial transaction history. Generative AI, on the other hand, has strengths in comprehensively reading unstructured text such as income verification, detailed bank deposit and withdrawal statements, text message payment information, and customer consultation records to derive decisions like loan approval or rejection.

However, from the perspective of financial institutions that must manage loan asset soundness, the text output of generative AI has acted as a serious limitation. Even if the AI outputs the word 'approve,' the delinquency risk of those borrowers is not the same, and a 'reject' decision also does not explain the absolute magnitude of risk. To operate key credit policies across the board—including loan approval, loan-to-value ratio calculation, limit setting, risk-based pricing, and loan-loss provisioning—a continuous default probability figure is essential. Until now, the word generation of LLMs has not been directly linked to such risk engines.

Upin's patent starts from the token log probabilities that the model calculates when predicting the next word. When the AI receives a review prompt and generates tokens such as approve, reject, or review, the log probabilities are extracted and converted into a rejection probability. The key is that this figure is not directly connected to the borrower's actual delinquency risk. The probability that the AI generates a specific word is merely a pre-trained linguistic statistic, and is different in nature from the likelihood of default arising under specific macroeconomic conditions or financial product environments.

To bridge this gap, the company designed a method of training a separate calibration function based on actual delinquency data of past borrowers. It clearly separates the response generation confidence of the language model from actual default rates in the financial market, and builds a conversion system that connects the two indicators through measured delinquency statistics. As a result, the review output of generative AI does not remain a simple judgment text but is transformed into a quantitative default probability acceptable to financial institutions, and can be directly linked to setting approval thresholds and applying differentiated interest rates according to policy standards.

This design makes a clear difference in risk management efficiency. When delinquency rate trends change sharply due to base rate fluctuations or economic recession, or when handling new financial product lines, retraining an LLM with tens of billions of parameters incurs astronomical computational costs and risks distorting review standards. In contrast, by keeping the AI model unchanged and retraining only a lightweight calibration function with delinquency measurements by country and product line, capital soundness indicators can be quickly recalibrated even amid rapid market changes. This is why it is evaluated as a safety mechanism that can immediately respond to signs of deterioration in local loan assets in emerging markets such as India, where mid- to low-credit borrowers with insufficient financial history are the majority.

However, for this calibration system to guarantee the long-term soundness of actual loan portfolios, several challenges must be solved. The reliability of the calibration function depends entirely on the quantity and precision of accumulated historical delinquency data. In new markets or vulnerable borrower segments where data is insufficient, estimation errors can increase. Also, if sudden collective delinquency occurs during an economic downturn that deviates from past patterns, there is a lag risk that the calibration function reflects the market shock belatedly. How to institutionally prove the transparency and explainability of default rate calculation grounds required by regulators for financial AI models in the commercialization process will be a key point determining success or failure.

※ This article is an AI translation of the Korean original published by Korea Credit News; it was checked by automated quality guards and published by the English edition desk.