Deep Learning-Based System and Method for Estimating Individual Interest Rate Sensitivity and Determining Optimal Loan Interest Rates

Abstract

The present invention relates to a system and method for determining an individual optimal loan interest rate by combining causal inference and game theory. Through a Double ML-based residual extraction engine, selection bias in interest rate determination can be eliminated, and the borrower's time discount rate (β) and delinquency penalty sensitivity (κ) can be inversely estimated as latent variables. Economic rationality is secured by internalizing game-theoretic equilibrium conditions into a neural network, and the imbalance problem of sparse delinquency data can be resolved through stepwise adaptive equilibrium sampling. By calculating an individual optimal interest rate that maximizes the financial institution's expected return based on the estimated interest rate sensitivity and applying it to real-time lending policies, profitability improvement and risk management can be achieved simultaneously.

Description

Technology Field

[0001] The present invention relates to artificial intelligence-based financial loan management technology, and more specifically, to a deep learning-based system and method that combines causal inference and game theory to estimate the interest rate sensitivity of individual borrowers and, based on this, determines an individual optimal loan interest rate that maximizes the expected return of a financial institution. More specifically, the invention relates to a technology that enhances risk-based pricing by eliminating selection bias in interest rate determination through a Double Machine Learning technique, inversely estimating potential economic propensities such as the borrower's time discount rate and delinquency penalty sensitivity, and predicting the borrower's strategic behavioral response by internalizing an economic equilibrium model into a neural network structure. Background Technology

[0002] In the lending operations of financial institutions, determining loan interest rates is a critical decision-making process that simultaneously influences profitability and risk management. Traditionally, financial institutions have adopted a method of predicting borrowers' default probabilities through credit rating models and applying uniform interest rates based on credit ratings. However, with the spread of digital finance and the emergence of fintech companies, the demand for personalized financial services is surging, and advancements in machine learning and artificial intelligence technologies are enabling more sophisticated risk assessment and pricing.

[0004] Existing credit-rating-based interest rate determination methods have a structural limitation in that they classify borrowers into groups based on their credit ratings and apply the same rate to all. This approach fails to reflect individual characteristics, such as interest rate sensitivity, repayment willingness, and economic circumstances, even within the same credit rating group, despite these differences. Consequently, there is a concern that inefficiencies may arise, such as charging excessive interest rates to prime borrowers to drive them away to competitors, while applying low rates to high-risk borrowers to undermine profitability.

[0006] Furthermore, existing machine learning-based delinquency prediction models are limited to predictions based on correlation and have the limitation of failing to estimate the causal effect of interest rate changes on borrower behavior. For example, while it is observed that borrowers with high interest rates exhibit high delinquency rates, there is a reverse causality problem in that it is impossible to distinguish whether the high interest rate was adopted due to high risk or whether the high interest rate itself caused the delinquency. Without the elimination of such selection bias, it is difficult to accurately predict the actual impact of interest rate adjustments on borrowers' loan acceptance probabilities or repayment behavior.

[0008] Furthermore, conventional predictive models have limitations in that they assume borrowers are passive agents and fail to consider their strategic responses to changes in interest rates. In reality, borrowers are active decision-makers who strategically decide whether to accept a loan and repay it in order to maximize their utility. Unobserved latent economic propensities, such as borrowers' time preference, risk appetite, and sensitivity to non-monetary costs associated with delinquency, are key factors determining responsiveness to interest rates, but existing systems face difficulties in estimating or utilizing these latent variables.

[0010] Furthermore, existing interest rate determination systems rely on static policies and have limitations in that they fail to reflect changes in the market environment or borrower behavioral feedback in real time. Despite the fact that dynamic factors—such as macroeconomic conditions, competitors' interest rate policies, and borrower income volatility—significantly impact loan risk and profitability, there is a problem in that there is a lack of mechanisms to adaptively reflect these factors.

[0012] Furthermore, due to the sensitivity of financial data, personal information protection regulations are being strengthened, and the demand for explainability in AI-based decision-making is increasing. However, existing black-box machine learning models struggle to transparently explain the rationale behind interest rate decisions, leading to difficulties in regulatory compliance and securing customer trust.

[0014] Against this backdrop, there is a need for a new technical approach that estimates the pure treatment effect of interest rates through causal inference techniques, predicts borrowers' strategic behavior using game-theoretic equilibrium models, and precisely calculates individual interest rate sensitivity by back-estimating unobserved potential economic parameters, and based on this, determines the optimal loan interest rate that maximizes the financial institution's expected returns.

[0016] The present invention can resolve the fundamental limitations of the existing uniform loan interest rate determination method based on credit ratings.

[0018] Conventional loan interest rate determination systems adopt a method of grouping borrowers by credit rating and applying the same interest rate, which may result in a failure to reflect differences in individual borrowers' interest rate sensitivity and repayment ability. From the perspective of financial institutions, this uniform approach can lead to customer attrition by imposing excessive interest rates on prime borrowers, or reduced profitability by applying low interest rates to high-risk borrowers.

[0020] Furthermore, existing systems are limited to merely predicting delinquency probabilities and may have the limitation of failing to estimate the causal effects of interest rate changes on borrower behavior. This implies that while they observe a correlation such as "borrowers with high interest rates have high delinquency rates," they may harbor a reverse causality problem, unable to distinguish whether borrowers accepted high rates due to high risk or whether delinquency occurred because of the high interest rates. Consequently, the inability to quantitatively predict the effects of policy interventions such as interest rate adjustments may lead to a problem where determining the optimal interest rate becomes impossible.

[0022] Furthermore, conventional technology may have limitations in that it utilizes only observable characteristics of borrowers (credit score, income, debt ratio, etc.) and fails to consider latent variables, such as the time discount rate or sensitivity to delinquency penalties, which represent the borrower's inherent economic propensity. Although these latent variables significantly influence a borrower's decision-making, they cannot be directly observed, which may make it difficult for existing systems to accurately estimate the borrower's interest rate sensitivity.

[0024] Furthermore, as delinquency is a rare event occurring in only a fraction of total loans, general machine learning models may suffer from a class imbalance problem where they are biased toward the majority class of timely repayments and fail to properly learn delinquency patterns. This can reduce the accuracy of delinquency probability predictions, thereby undermining the reliability of risk management. Prior art literature

[0025] Korean Registered Patent KR 10-2528480 B1 Korean Registered Patent KR 10-2565223 B1 The problem to be solved

[0026] Accordingly, the present invention may aim to provide a system and method for determining an optimal loan interest rate that maximizes the expected return of a financial institution by estimating the pure causal effect of interest rates through causal inference techniques, inversely calculating the borrower's potential economic parameters using a game-theoretic equilibrium model, and precisely calculating individual interest rate sensitivity based on this. means of solving the problem

[0027] A deep learning-based individual interest rate sensitivity estimation and optimal loan interest rate determination system using the same according to one aspect of the present invention for solving the above problem may include: a data preprocessing unit that synchronizes the borrower's credit information and financial behavior data based on T1, which is the time of loan approval; a Double ML-based residual extraction engine that extracts prediction residuals from a first learning model that predicts a delinquency probability (y) based on borrower characteristics (X) and a second learning model that predicts the interest rate (T) proposed by the financial company, respectively; a latent parameter estimation unit that estimates the time discount rate (β) and delinquency penalty sensitivity (κ), which are unobserved implicit economic propensities of the borrower, as latent variables based on the extracted residuals; and an optimization control unit that calculates an individual optimal loan interest rate that maximizes the financial company's expected profit based on the estimated latent variables and the individual interest rate sensitivity calculated based on a game theory equilibrium model.

[0029] The above residual extraction engine can perform balanced sampling to adjust the ratio of normal samples to delinquent samples to a preset ratio in order to correct the sparsity of delinquent data during the training of the first learning model and the second learning model.

[0031] The above latent parameter estimator can inversely learn the time discount rate (β) and delinquency penalty sensitivity (κ) by including an economic constraint in the loss function that the probability of a loan execution decreases monotonically as interest rates rise.

[0033] The above data preprocessing unit may include a behavior-based propensity extractor that quantifies the borrower's risk-taking propensity by analyzing past loan application cycles and repayment behavior data, rather than the borrower's survey data.

[0035] The optimization control unit above can use a full dataset that was not used in the learning phase of the first learning model and the second learning model as input to the residual regression model for calculating the optimal loan interest rate.

[0037] A method for estimating individual interest rate sensitivity based on deep learning and determining an optimal loan interest rate using the same, according to another aspect of the present invention for solving the above problem, may include the steps of: collecting financial behavior data of a borrower and synchronizing a time series based on the time of loan approval; applying equilibrium sampling to the synchronized data to extract the borrower's delinquency residual and interest rate treatment residual, respectively; estimating the borrower's potential economic parameters (β, κ) and interest rate sensitivity from the extracted residuals through an artificial neural network containing game-theoretic equilibrium conditions; and determining an individual optimal loan interest rate by searching for the maximum point of an expected return function with the estimated interest rate sensitivity as a variable.

[0039] The step of estimating the above parameters and interest rate sensitivity can be performed by integrating an economic decision-making function, in which a borrower decides whether to accept and repay a loan to maximize their utility, into the activation function of the above artificial neural network.

[0041] The above method may further include the step of executing a real-time loan policy by transmitting the determined optimal loan interest rate data to an external credit screening engine via an API (Application Programming Interface).

[0043] A latent parameter inversion estimation engine according to another aspect of the present invention for solving the above problem may include a module that receives observable borrower loan execution and repayment data, and an artificial neural network module that defines the borrower's time preference (β) and default cost sensitivity (κ) as latent variables and inversely calculates the optimal latent variable value that caused the input data based on economic equilibrium conditions.

[0045] A causal inference-based price optimization engine according to another aspect of the present invention for solving the above problem may include a module that generates residual data from which selection bias in interest rate determination is removed through a Double ML technique, and a price optimization engine that combines the residual data with parameters derived from a latent parameter inversion estimation engine to causally predict changes in borrower behavior due to interest rate fluctuations and calculate an interest rate value that maximizes expected returns. Effects of the invention

[0046] According to the present invention, by eliminating selection bias in interest rate decisions and estimating the pure causal effect of interest rate changes through a Double ML-based causal inference framework, it is possible to provide a significant technical effect that can quantitatively predict the effect of policy intervention compared to conventional correlation-based prediction methods. In particular, due to a mechanism that minimizes the influence of unobserved confounding variables by extracting the residuals of the first learning model and the second learning model using a cross-validation method, it is possible to solve the technical challenge of the reverse causality problem that could not be solved by existing technologies.

[0048] In addition, the technical feature of the present invention of incorporating the game-theoretic equilibrium model into a neural network structure enables the inverse estimation of unobservable latent variables, such as the borrower's time discount rate (β) and delinquency penalty sensitivity (κ), thereby providing an advantageous effect of precisely identifying the borrower's intrinsic economic propensity compared to conventional methods that utilize only observable characteristics. This can be achieved through a specific implementation method that integrates an economic utility maximization function into a loss function and imposes monotonicity constraints.

[0050] Furthermore, the balanced sampling technique of the present invention can provide a technical effect that significantly improves the degradation of prediction accuracy caused by the conventional class imbalance problem by sufficiently learning the pattern of the rare event known as delinquency while utilizing all data in the final prediction stage to reflect the actual distribution.

[0052] Furthermore, the behavior-based propensity extraction technology of the present invention has the advantageous effect of enabling objective and highly reliable propensity measurement by eliminating response bias and social desirability bias through the quantification of risk-taking propensity from actual financial behavior data rather than surveys.

[0054] According to the present invention, by precisely calculating individual interest rate sensitivity and determining a customized optimal interest rate, the expected return of a financial institution can be significantly improved compared to the conventional uniform interest rate method based on credit rating. This can provide a dual effect of preventing customer attrition caused by the imposition of excessive interest rates on prime borrowers and minimizing profit loss caused by the application of low interest rates to high-risk borrowers.

[0056] In addition, the causal effect-based risk prediction mechanism of the present invention enables the quantitative evaluation of the impact of interest rate adjustments on the probability of delinquency in advance, thereby providing a significant technical effect that enables preemptive and sophisticated risk management compared to conventional ex-post risk management methods.

[0058] Furthermore, the present invention has the advantageous effect of overcoming the limitations of black-box AI through a system structure interpretable in terms of economic parameters, facilitating financial regulatory compliance and audit response, and automatically verifying non-discrimination principles.

[0059] Meanwhile, the effects obtainable from the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art to which the present invention belongs from the description below. Brief explanation of the drawing

[0060] The following drawings attached to this specification illustrate preferred embodiments of the present invention and serve to further enhance understanding of the technical concept of the present invention together with the detailed description of the invention; therefore, the present invention should not be interpreted as being limited only to the matters described in such drawings. FIG. 1 is an overall block diagram of the MISTnet system according to the present invention, illustrating a system architecture including a multidimensional financial behavior data preprocessing and adaptive sampling module (100), a latent parameter estimation and Double ML engine (200), a decision modeling module with embedded game theory equilibrium conditions (300), and an optimal interest rate determination and control API (400) based on expected return maximization. FIG. 2 is a flowchart illustrating the overall process of the MISTnet-based method for estimating individual interest rate sensitivity and calculating the optimal interest rate according to the present invention, and illustrates an operation process including a step of generating a dataset for causal inference and time synchronization (S310), a step of calculating residuals based on Double ML first and second stage balanced sampling (S320), a step of inverting latent parameters and estimating causal effects based on Double ML third stage all data (S330), a step of determining the optimal loan interest rate based on an expected return function (S340), and a step of linking with a loan approval engine and real-time control (S350). Figure 3 is a detailed structural diagram of the Double ML-based MISTnet 3-stage architecture, which is the core of the present invention, illustrating a pipeline in which a first learning model and a second learning model calculate delinquency residuals and interest rate treatment residuals in parallel, and combine these residuals in the third stage to estimate causal effects. Figure 4 is a neural network layer configuration diagram for estimating latent parameters (β, κ), illustrating an artificial neural network structure that outputs acceptance probability and delinquency probability by inversely calculating the borrower's time discount rate (β) and delinquency penalty sensitivity (κ) in the latent variable layer. Figure 5 is a graph showing a decision probability curve reflecting game theory equilibrium conditions, illustrating the result of implementing economic rationality through MaxLoss and MISTLoss loss functions, where the probability of execution decreases monotonically and the probability of delinquency increases as interest rates rise. Figure 6 is a diagram for calculating the Sweet Spot for determining the optimal interest rate to maximize expected returns, illustrating the process of determining the optimal interest rate for each individual by exploring the maximum point of the expected return function with the estimated potential parameters as variables. Figure 7 is a comparative diagram of step-by-step sampling strategies for solving data sparsity, illustrating balanced sampling that adjusts normal and overdue samples in a 10:1 ratio during the first and second stages of learning, and adaptive sampling strategy that utilizes all data during the third stage of causal inference. Specific details for implementing the invention

[0061] The above objects, other objects, features, and advantages of the present invention will be easily understood through the following preferred embodiments associated with the accompanying drawings. However, the present invention is not limited to the embodiments described herein and may be embodied in other forms. Rather, the embodiments introduced herein are provided to ensure that the disclosed content is thorough and complete and to ensure that the spirit of the invention is sufficiently conveyed to a person skilled in the art.

[0062] In this specification, when a component is described as being on another component, it means that it may be formed directly on the other component or that a third component may be interposed between them. Also, in the drawings, the thicknesses of the components are exaggerated for the effective description of the technical content.

[0063] The embodiments described herein will be explained with reference to cross-sectional and / or plan views, which are exemplary illustrations of the invention. In the drawings, the thicknesses of films and regions are exaggerated for effective explanation of the technical content. Accordingly, the shapes of the exemplary drawings may be modified by manufacturing techniques and / or tolerances, etc. Accordingly, the embodiments of the invention are not limited to the specific shapes depicted but include variations in shape produced according to the manufacturing process. For example, a region depicted as a right angle may be rounded or have a certain curvature. Accordingly, the regions illustrated in the drawings have properties, and the shapes of the regions illustrated in the drawings are intended to illustrate specific shapes of the regions of the device and are not intended to limit the scope of the invention. Although terms such as first, second, etc., have been used to describe various components in the various embodiments of this specification, these components should not be limited by such terms. These terms are used merely to distinguish one component from another. The embodiments described and illustrated herein also include their complementary embodiments.

[0064] The terms used herein are for describing the embodiments and are not intended to limit the invention. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used herein, 'comprises' and / or 'comprising' do not exclude the presence or addition of one or more other components to the mentioned components.

[0065] In describing the specific embodiments below, various specific details have been included to explain the invention more specifically and to aid understanding. However, a reader with sufficient knowledge in the art to understand the invention will recognize that it can be used without these various specific details. In some cases, it is noted in advance that commonly known aspects that are not significantly related to the invention have been omitted to prevent unnecessary confusion in describing the invention.

[0067] The present invention relates to a system and method for determining an optimal loan interest rate that maximizes the profit of a financial institution by precisely estimating the responsiveness of individual borrowers to changes in interest rates. Existing uniform interest rate determination methods based on credit ratings fail to consider the interest rate sensitivity of individual borrowers, which can lead to problems such as the loss of prime customers or reduced profitability due to the imposition of under-interest rates on high-risk customers. Furthermore, simple prediction models may find it difficult to estimate the pure causal effect of interest rate changes because they fail to eliminate selection bias that occurs during interest rate determination.

[0069] To solve these technical problems, the present invention can provide an innovative system that calculates an optimal loan interest rate by combining a Double Machine Learning (Double ML)-based causal inference framework with a game-theoretic equilibrium model to estimate the borrower's unobserved implied economic propensity as a latent variable and deriving individual interest rate sensitivity based thereon. The present invention can quantitatively measure the pure effect of interest rate changes through causal inference and simultaneously ensure prediction accuracy and interpretability by internalizing economic rationality into a neural network structure.

[0071] FIG. 1 is a block diagram illustrating the overall configuration of a deep learning-based individual interest rate sensitivity estimation and optimal loan interest rate determination system using the same according to an embodiment of the present invention. Referring to FIG. 1, the system of the present invention may include a multidimensional financial behavior data preprocessing and adaptive sampling module (100), a latent parameter estimation and Double ML engine (200), a decision modeling module with embedded game theory equilibrium conditions (300), and an optimal interest rate determination and control API (400) based on expected return maximization.

[0073] The multidimensional financial behavior data preprocessing and adaptive sampling module (100) can perform the role of integrating and refining heterogeneous financial data to increase the precision of causal inference. This module (100) may include a heterogeneous data integrator (110), a behavior-based implied propensity extractor (120), and a step-by-step adaptive balance sampler (130).

[0075] The heterogeneous data integrator (110) can collect and integrate data from various sources, such as the borrower's credit information agency (CB) data, financial transaction history, and behavior logs within the application. At this time, time series data segmented into loan inquiry time (T0), approval time (T1), execution time (T2), and delinquency occurrence time (T3) can be synchronized based on the loan approval time (T1) to form a structure that allows for causal relationship analysis. This time synchronization enables comparative analysis of each borrower's financial behavior on a consistent time axis, thereby increasing the learning efficiency of the model.

[0077] The behavior-based implied propensity extractor (120) can quantify the borrower's implied propensity from actual observable financial behavior data without relying on the borrower's survey data. Specifically, by analyzing the volatility of past loan application cycles, repayment behavior patterns, frequency of early repayment, speed of recovery after delinquency, and acceptance by interest rate range, the borrower's risk-taking propensity, impulsivity, financial planning ability, and sense of responsibility can be extracted as digital features. This is based on the theory of revealed preference, which utilizes the economic principle that the borrower's actual behavior reflects their implied preferences. For example, the value obtained by dividing the standard deviation of the loan application cycle by the mean can be used as an indicator of impulsivity, and the frequency of early repayment can be used as a measure of financial planning ability.

[0079] A stepwise adaptive balance sampler (130) may be a component for solving class imbalance problems specific to financial data. Generally, in loan data, delinquency is a sparse event occurring in only a small fraction of total loans; therefore, if this is used directly for training, the model may be biased toward normal repayments, which are the majority class. The stepwise adaptive balance sampler (130) may include a sparse data correction engine (121) and a full data pass filter (122).

[0081] The sparse data correction engine (121) can perform balanced sampling to adjust the ratio of normal samples to delinquent samples to a preset ratio during the learning phase of the first learning model and the second learning model. According to some embodiments, the ratio of normal samples to delinquent samples can be adjusted to a level of 10:1, but this may be set to a different ratio depending on the characteristics of the data and the requirements of the model. This balanced sampling can be implemented through techniques such as SMOTE (Synthetic Minority Over-sampling Technique), undersampling, or class weight adjustment. Through balanced sampling, the model can sufficiently learn sparse delinquent patterns and increase the sensitivity of causal relationships by assigning weights to samples near the decision boundary.

[0083] The full data pass filter (122) allows the entire dataset to be used as is without applying balanced sampling during the third stage of residual regression modeling. This may be intended to ensure the generalization performance of the model by reflecting the data distribution of the actual market in the final causal effect estimation stage, unlike the first and second learning stages for bias removal. This step-by-step differentiated sampling strategy is a unique data processing technique of the present invention, which can simultaneously achieve stability of learning and accuracy of prediction.

[0085] The latent parameter estimation and Double ML engine (200) may be a core computational unit that mathematically removes selection bias occurring during interest rate determination and estimates the borrower's unobserved implied economic propensity as a latent variable. This engine (200) may include a first learning unit (210), a second learning unit (220), and a latent variable inverse calculator (230).

[0087] The first learning unit (210) can implement a first learning model that predicts a delinquency probability (y) using borrower characteristics (X) as input. This model can be implemented using various machine learning algorithms, such as Gradient Boosting Machine, Random Forest, or a deep learning-based classifier. The first learning unit (210) can predict the probability that a borrower will be delinquent based on characteristics such as the borrower's credit score, income, debt ratio, and employment type, and can calculate a delinquency residual, which is the difference between the actual delinquency status and the predicted value. This residual may represent residual variability in the occurrence of delinquency that is not explained by the characteristics of the borrower.

[0089] The second learning unit (220) can implement a second learning model that predicts the interest rate (T) to be offered by a financial company using the same borrower characteristics (X) as input. This model can be implemented as a regression algorithm and can learn the financial company's interest rate determination policy. Generally, financial companies offer lower interest rates to borrowers with high credit ratings, so such policy bias may be inherent in the data. The second learning unit (220) can calculate an interest rate treatment residual, which is the difference between the predicted interest rate and the actual interest rate offered. This residual may represent residual volatility in interest rate determination that is not explained by the observable characteristics of the borrower.

[0091] The first training unit (210) and the second training unit (220) can generate residuals using a K-fold cross-validation method. Specifically, the entire data can be divided into K folds, and for each fold, a model can be trained using the data from the remaining folds, and then the residuals for that fold can be calculated. This prevents overfitting and generates residuals with bias removed. Additionally, the balanced sampling described earlier can be applied at this stage to improve the learning efficiency of sparse overdue data.

[0093] The latent variable inverse calculator (230) takes the residuals extracted from the first learning unit (210) and the second learning unit (220) as input, and the time discount rate of the next week ( β ) and delinquency penalty sensitivity( k ) can be estimated as a latent variable. Time discount rate( β ) is a parameter representing the extent to which the borrower discounts future value to present value, and can reflect the borrower's patience or time preference. β The higher the value, the less important borrowers may consider future repayment burdens to be compared to current consumption utility, which can increase the probability of loan acceptance and raise the likelihood of delinquency. Delinquency penalty sensitivity ( k ) may indicate sensitivity to non-monetary costs (credit score decline, social stigma, psychological stress, etc.) that borrowers experience when delinquency occurs. k The higher the value, the stronger the motivation for the borrower to avoid delinquency.

[0095] The latent variable inverse calculator (230) can be implemented as an artificial neural network structure, the input layer may include delinquency residuals, interest rate treatment residuals, and behavioral propensity indicators, the hidden layer may be composed of a fully connected layer, and the output layer β and k It can output estimated values. This neural network can be trained using the Inverse Optimization method. That is, it can explore the optimal latent variable values ​​that best explain the observed borrower behavior (loan acceptance / rejection, repayment / delinquency). To this end, the borrower's utility function can be defined. For example, the utility upon loan acceptance is U (Acceptance) = Loan Amount β × Future repayment amount k It can be expressed as an expected delinquency penalty, and the utility upon rejection is U (Rejection) can be set to 0. It can be assumed that the borrower makes decisions in a direction that maximizes this utility, and the acceptance probability is P (Accept)= s ( U (acceptance) U It can be modeled as (rejection). Here s can be a sigmoid function.

[0097] The loss function of the latent variable inverse generator (230) can be designed to minimize the difference between the actual observed behavior and the behavior predicted by the model. For example, a cross-entropy loss can be used, and economic constraints can be added thereto. In particular, a constraint that the probability of a loan execution monotonically decreases as interest rates rise can be included as a penalty term in the loss function. This is Penalty= l ×∑max(0, P ( r high) P ( r It can be expressed in the form of low), where l can be a hyperparameter that controls the strength of the constraint. Such monotonicity constraints can ensure economic rationality and prevent illogical results caused by data noise.

[0099] The decision-making modeling module (300) with embedded game theory equilibrium conditions can model the strategic interaction between the borrower and the financial institution from a game theory perspective and embed economic rationality into the neural network structure. This module (300) may include a game theory equilibrium constraint condition (310) and a MaxLoss & MISTLoss optimizer (320).

[0101] The game theory equilibrium constraint condition (310) can directly incorporate the economic assumption that the borrower maximizes their utility into the activation function or layer of the neural network. Specifically, the borrower's decision function can be defined to model the choice of accepting / rejecting a loan and repaying / delinquently. This can be understood as applying the concept of a Physics-Informed Neural Network to economics. It can be assumed that the borrower makes a choice that maximizes their utility under given interest rate conditions, and the model can be trained so that the solution to this optimization problem matches the observed behavior.

[0103] The game theory equilibrium constraint (310) can utilize the concept of Nash Equilibrium. The borrower and the financial company each seek to maximize their own interests and can select an optimal strategy by considering the other party's strategy. The financial company predicts the borrower's interest rate sensitivity and proposes an optimal interest rate, and the borrower can decide whether to accept the proposed rate. Finding the equilibrium point of this interaction may be the role of the game theory equilibrium constraint (310).

[0105] The MaxLoss & MISTLoss optimizer (320) can apply specialized loss functions to prevent the reversal of causal relationships and ensure economic validity. MaxLoss may include a monotonicity constraint to control the loan execution probability so that it does not increase illogically when interest rates rise. MISTLoss is short for Monotonicity-Informed Structural Training Loss and may be a loss function that incorporates economic structural constraints into the training process. In addition to minimizing general prediction errors, this loss function may impose a penalty to ensure that the relationship between interest rates and execution probability remains within an economically valid range.

[0107] According to some embodiments, the MaxLoss & MISTLoss optimizer (320) may perform constraint-conditional optimization using the Lagrange Multiplier Method. Alternatively, the constraints may be converted into soft constraints and added as weighted penalty terms to the loss function. Through this, the model can find a balance between data fit and economic feasibility.

[0109] The API (400) for determining and controlling the optimal interest rate based on expected return maximization can calculate an individual optimal loan interest rate that maximizes the financial company's expected return based on estimated potential parameters and interest rate sensitivity, and apply this to the lending system in real time. This API (400) may include an expected return function generator (410), a return maximization interest rate calculator (420), and a real-time lending policy linkage interface (430).

[0111] The expected return function generator (410) is estimated β and k Based on this, the financial firm's expected return curve based on interest rate changes can be generated. The expected return function can be defined as follows:

[0113] E [Profit| r ]= P (Accept| r )×[ r ×Principal× P (Repayment|Accept, r )Principal × (1 P (Repayment|Accept, r ))×Loss Rate]

[0115] Here r Silver interest rate, P (Accept| r ) is interest rate r Given this, the probability that the borrower will accept the loan, P (Repayment|Accept, r ) can represent the probability that a borrower who has accepted a loan will repay it. P (Accept|r ) can be calculated from a utility function utilizing estimated latent parameters, and P (Repayment|Accept, r ) can be derived by reflecting interest rate effects in the predicted value of the first learning model. The loss rate may represent the proportion of unrecoverable amounts when delinquency occurs.

[0117] The expected return function generator (410) can generate a unique expected return curve for each individual borrower. This curve can generally have an upwardly convex shape, and if the interest rate is too low, interest income is insufficient, and if the interest rate is too high, the borrower may refuse the loan or the probability of default increases, which may reduce income.

[0119] The profit maximization interest rate calculator (420) can calculate an individual optimal loan interest rate by searching for the maximum point of the generated expected profit function. This can be performed through a numerical optimization algorithm, for example, the L-BFGS-B (Limited-memory Broyden-Fletcher-Goldfarb-Shanno with Bound constraints) algorithm may be used. This algorithm is suitable for constraint-conditional optimization and can take into account constraints such as the legal maximum interest rate ceiling, the financial institution's internal risk policy, and the interest rate floor to prevent customer churn.

[0121] The optimization problem can be formulated as follows:

[0123] r max E [Profit| r ]

[0124] subject to r min≤ r ≤ r max

[0126] Here r min and rmax may represent the lower and upper limits of the interest rate, respectively. According to some embodiments, additional constraints may be included. For example, portfolio-level risk constraints, constraints on interest rate dispersion across specific credit rating ranges, or constraints on maintaining competitiveness against competitor rates may be considered.

[0128] The profit maximization interest rate calculator (420) can perform gradient-based optimization using a differentiable expected profit function. Alternatively, techniques such as grid search or Bayesian optimization may be used. The optimal interest rate may be a customized value that reflects the characteristics of the individual borrower, estimated potential parameters, and current market conditions.

[0130] The real-time loan policy linkage interface (430) can execute a real-time loan policy by transmitting the calculated optimal interest rate to an external loan screening engine. This interface (430) can be implemented as a RESTful API or a message queue and can transmit data in JSON or XML format. The transmitted data may include a borrower identifier, optimal interest rate, estimated delinquency probability, interest rate sensitivity, and confidence interval.

[0132] The real-time loan policy linkage interface (430) can support bidirectional communication. That is, it can receive loan execution results and repayment history data from the loan screening engine and use them for continuous learning and updating of the model. Through this, the system can dynamically adapt to market changes and changes in borrower behavior patterns.

[0134] According to some embodiments, the real-time credit policy linkage interface (430) may integrate an A / B test framework. Performance can be compared and evaluated by applying the existing interest rate policy to some borrowers and the optimal interest rate calculated by the system of the present invention to others. This allows for the verification of the actual effectiveness of the system and the gradual expansion of its scope of application.

[0136] FIG. 2 is a flowchart illustrating a method for estimating individual interest rate sensitivity based on deep learning and determining an optimal loan interest rate using the same, according to an embodiment of the present invention. Referring to FIG. 2, the method (S300) of the present invention may include a step of generating a dataset for causal inference and time synchronization (S310), Double ML steps 1 and 2: a step of calculating residuals based on balanced sampling (S320), Double ML step 3: a step of inverting latent parameters based on all data and estimating causal effects (S330), a step of determining an optimal loan interest rate based on an expected return function (S340), and a step of linking with a loan approval engine and real-time control (S350).

[0138] In the step of generating a dataset for causal inference and time synchronization (S310), financial behavior data of borrowers can be collected, and time series can be synchronized based on the loan approval time. Specifically, heterogeneous data segmented at the loan inquiry time (T0), approval time (T1), execution time (T2), and delinquency occurrence time (T3) can be aligned based on the loan approval time (T1). Through this, each borrower's past financial history can be transformed into a relative time axis prior to T1 (e.g., T-1 month, T-2 month), and future repayment history into a relative time axis after T1 (e.g., T+1 month, T+2 month).

[0140] In addition, behavior-based propensity extraction can be performed at this stage. Instead of borrower survey data, the borrower's inherent propensity can be quantified by extracting revealed preference data—such as actual past loan application cycles, repayment lead times, and acceptance rates by interest rate range—as digital features. For example, volatility in loan application cycles can be used as an indicator of impulsivity, the frequency of early repayments as an indicator of financial planning, and the speed of recovery after delinquency as an indicator of responsibility.

[0142] Double ML Steps 1 and 2: The balanced sampling-based residual calculation step (S320) may be a step of training two independent models in parallel to mathematically eliminate selection bias in interest rate determination. This step may include a first training and delinquency residual derivation step (S321) and a second training and interest rate residual derivation step (S322).

[0144] In the first learning and deduction of delinquency residual step (S321), a first learning model that predicts delinquency status (y) can be trained using the characteristics (X) of the borrower as input. At this time, balanced sampling can be applied to overcome the sparsity of delinquency data by adjusting the ratio of normal samples to delinquency samples to a preset ratio. According to some embodiments, the ratio of normal to delinquent samples can be adjusted to a level of 10:1, but this may be set to a different ratio depending on the characteristics of the data. The probability of delinquency is predicted through the trained model, and the delinquency residual (which is the difference between the actual value and the predicted value) e y ) can be produced.

[0146] In the second learning and interest rate residual derivation step (S322), a second learning model can be trained to predict the interest rate (T) to be proposed by the financial institution using the same borrower characteristics (X) as input. By training using balanced sampling, the bias-removed interest rate treatment residual ( e T )These two residuals can be calculated. These two residuals represent volatility that cannot be explained by the observable characteristics of the borrower, and the pure causal effect of the interest rate can be estimated by analyzing the correlation between them.

[0148] The first and second training can be performed using the K-fold cross-validation method. The entire dataset is divided into K folds, and for each fold, a model is trained using the data from the remaining folds, after which the residuals for that fold can be calculated. This prevents overfitting and generates residuals with bias removed.

[0150] Double ML Step 3: The step of inverse calculation of latent parameters and causal effect estimation based on total data (S330) is a core step of the present invention, wherein the latent economic parameters of the borrower ( β , k ) and interest rate sensitivity can be estimated.

[0152] In this stage, unlike stages 1 and 2, the generalization performance of the model can be maximized by inputting the entire dataset that has not been balanced or sampled. Utilizing the entire dataset can improve the accuracy of the final prediction by reflecting the data distribution of the actual market.

[0154] In the back-calculation process of latent variables, the time discount rate hidden within the borrower ( β ) and delinquency penalty sensitivity( k ) can be defined as a latent variable. Time discount rate( β ) indicates the borrower's patience or time preference, and the late payment penalty sensitivity( k ) can indicate sensitivity to psychological bankruptcy costs.

[0156] By applying game-theoretic constraints, the assumption of economic rationality—that borrowers maximize their utility—can be incorporated as a constraint into the neural network learning process. Specifically, the borrower's utility function can be defined, and the model can be trained so that the choice maximizing it matches the observed behavior. Additionally, economic validity can be ensured by including the monotonicity constraint, which states that the probability of a loan execution decreases when interest rates rise, in the loss function.

[0158] The interest rate sensitivity of individual borrowers can be derived as the output of this stage. Interest rate sensitivity can be an indicator that quantifies the degree to which an individual borrower's default probability or loan acceptance probability changes when interest rates fluctuate by a predetermined level. This can be expressed as the concept of elasticity, and e =Δ r / r Δ P / P It can be calculated in the form. Here P is the probability of a loan being executed, r Silver can represent interest rates.

[0160] In the optimal loan interest rate determination step based on the expected return function (S340), the optimal loan interest rate for each individual can be determined by exploring the maximum point of the expected return function with the estimated interest rate sensitivity as a variable.

[0162] First, a profit curve can be generated. Based on the derived interest rate sensitivity and latent parameters, a profit function for each individual customer can be generated by combining the financial institution's expected interest income and expected default loss in response to changes in interest rates. This function may be one in which the interest rate is the independent variable and expected profit is the dependent variable.

[0164] Next, the optimal point can be calculated. By differentiating the generated profit function or applying numerical optimization algorithms, the final individual optimal loan interest rate can be determined, which maximizes profit while adhering to the risk tolerance range. This optimal rate may be an equilibrium point that considers both the borrower's churn potential and delinquency risk.

[0166] During the optimization process, constraints such as the legal maximum interest rate ceiling, the financial institution's internal risk policies, and interest rate floors designed to prevent customer churn may be considered. Additionally, portfolio-level risk constraints or restrictions on maintaining competitiveness against competitors' rates may be added.

[0168] In the loan approval engine integration and real-time control step (S350), the determined optimal loan interest rate data can be transmitted to an external loan screening engine via API to execute a real-time loan policy.

[0170] During the API transmission process, data such as borrower identifiers, optimal interest rates, estimated delinquency probabilities, interest rate sensitivity, and confidence intervals can be transmitted in JSON or XML format. The transmitted data can be immediately utilized by the credit screening system to present customized loan offers to borrowers.

[0172] Through the application of dynamic policies, financial institutions can execute customized limit and interest rate offers for individual customers in real time based on the calculated interest rate sensitivity profile. Additionally, they can receive loan execution results and repayment history data to utilize for the continuous learning and updating of the model.

[0174] According to some embodiments, A / B testing may be performed at this stage. Performance can be compared and evaluated by applying the existing interest rate policy to some borrowers and the optimal interest rate calculated by the system of the present invention to others. This allows for the verification of the actual effectiveness of the system and the gradual expansion of its scope of application.

[0176] Figure 3 illustrates the detailed structure of the Double ML-based three-stage architecture of the present invention. Referring to Figure 3, the causal inference pipeline is visualized, which clearly demonstrates the difference from a simple classification model.

[0178] A parallel block structure can be illustrated at the top, where borrower characteristics (X) are input and branch into a first learning model (delinquency prediction) and a second learning model (interest rate prediction), respectively. At the output of each model, the delinquency residual, which is the difference from the actual value, ( e y ) and interest rate treatment residuals( e T The process of producing ) can be expressed.

[0180] At the bottom, the flow in which the two calculated residuals are integrated and enter the third-stage causal effect estimation engine can be connected by an arrow. This structure is intended to mathematically eliminate selection bias that occurs during interest rate determination; it offsets data distortion caused by the tendency of financial institutions to offer lower interest rates to prime customers through the extraction of residuals in the first and second stages, and allows for the derivation of only the pure responsiveness to changes in interest rates in the third stage.

[0182] Figure 4 shows a latent parameter ( β , k This illustrates the configuration of neural network layers for estimation. Referring to Fig. 4, the layer structure of the artificial neural network is shown vertically, and a latent parameter layer can be placed in the middle as a separate highlighted block.

[0184] This layer includes the car owner's time discount rate ( β ) and delinquency penalty sensitivity( k ) can be defined and displayed as a latent variable. The final output is the loan acceptance probability ( P accept) and delinquency probability ( P It can be a structure connected via delinquency.

[0186] This system can utilize not only surface data such as the borrower's credit score, but also the borrower's intrinsic economic parameters, which are inversely calculated by artificial intelligence through past behavioral data, as key learning factors. β The car owner's patience, k This refers to the psychological default cost, which allows for the prediction of the borrower's future behavior using game theory.

[0188] The input layer of the neural network may include delinquency residuals, interest rate treatment residuals, and behavioral propensity indicators, and the hidden layer may consist of a fully connected layer. The Rectified Linear Unit (ReLU) may be used as the activation function, and in the output layer β For the Sigmoid function, k You can apply the Softplus function to each to output values ​​in the range [0,1] and positive ranges, respectively.

[0190] Figure 5 illustrates a decision probability curve reflecting game theory equilibrium conditions. Referring to Figure 5, a graph can be illustrated where the X-axis represents the interest rate and the Y-axis represents the probability. A logical curve can be represented in which the probability of loan execution (acceptance) smoothly decreases as the interest rate rises, while the probability of delinquency increases.

[0192] At the bottom of the graph, a block may be displayed indicating that the loss function includes game theory constraints. This may be a visualization of constraints that prevent illogical probability calculations.

[0194] Simple statistical models can produce illogical results where acceptance probabilities increase even as interest rates rise due to data noise. This invention directly integrates a game-theoretic equilibrium model into the loss functions (MaxLoss, MISTLoss) to enforce that probabilities are calculated only within an economically valid range. This ensures that the model's predictions adhere to economic rationality.

[0196] Figure 6 illustrates a diagram for calculating the Sweet Spot for determining the optimal interest rate to maximize expected returns. Referring to Figure 6, an upwardly convex parabolic curve can be depicted, with the X-axis representing the interest rate and the Y-axis representing the expected return. An indicator line representing the optimal interest rate (Sweet Spot) can be displayed at the apex of the curve.

[0198] This curve is estimated β and k It can be derived from an expected return function with variables. Based on the estimated individual interest rate sensitivity, an individualized optimal interest rate can be calculated that maximizes the financial institution's profit while minimizing customer churn and default losses.

[0200] Unlike simple risk classification methods, the present invention may be a profit optimization engine that integrates pricing and risk management. If interest rates are too low, interest income is insufficient, and if interest rates are too high, profits may decrease as borrowers may refuse loans or the probability of delinquency increases. The optimal interest rate may be an equilibrium point that takes these trade-offs into account.

[0202] Figure 7 illustrates a comparison of step-by-step sampling strategies for resolving data sparsity. Referring to Figure 7, on the left, a balanced sampling dataset, in which normal data and sparse overdue data are matched at a preset ratio for the first and second stages of training, can be represented as a block. On the right, a structure in which an unsampled total dataset is input as is for the third stage of causal inference can be illustrated.

[0204] A dual-structure data pipeline can be described to overcome the chronic problem of insufficient delinquency samples in financial data. An adaptive sampling strategy that uses balanced sampling in the first and second stages for bias removal and full data in the third stage for actual effect calculation may be a unique data processing technique of the present invention.

[0206] This strategy can achieve both learning stability and prediction accuracy simultaneously. In the first and second stages, the model can sufficiently learn sparse delinquency patterns through balanced sampling, and in the third stage, generalization performance can be secured by using full data to reflect the actual market distribution.

[0208] This embodiment can explain the process of deriving individual interest rate sensitivity while solving the scarcity problem unique to financial data.

[0210] First, during the data preparation and synchronization phase, the system collects the borrower's credit information and behavior logs within the application, and can align all features based on T1, the time of loan approval. This allows each borrower's financial behavior to be compared and analyzed over a consistent time axis.

[0212] Next, first and second-stage balanced sampling learning can be performed. Considering that the delinquency rate among all loan executions is low, the system can perform balanced sampling by adjusting the ratio of normal samples to delinquent samples to a preset ratio when predicting delinquency status (first learning) and interest rate (second learning). Through this, the bias-removed baseline delinquency residual ( ey ) and interest rate treatment residuals( eT ) can be produced.

[0214] In the third stage of total data causality estimation, the generalization performance of the model can be secured by inputting an unsampled total dataset of the actual market into the regression model that estimates causal effects on the extracted residuals.

[0216] As an effect of this embodiment, by blocking data distortions biased toward specific interest rate ranges, it is possible to accurately derive the pure causal sensitivity regarding how the execution probability and delinquency probability of individual borrowers change when interest rates fluctuate.

[0218] This embodiment can explain the process by which an artificial neural network infers the unobserved intrinsic economic propensity of a vehicle owner.

[0220] First, in the latent variable definition phase, the system uses a time discount rate representing the borrower's patience ( β ) and sensitivity to delinquency penalties, which is the psychological distress felt upon default ( k ) can be set as a latent variable layer inside the neural network.

[0222] Next, in the game theory equilibrium injection phase, it can be assumed that the borrower selects the loan acceptance and repayment paths based on economic rationality, which assumes maximizing their utility. The system can directly incorporate the monotonicity constraint—that the probability of a loan execution decreases as interest rates rise—into the neural network's activation function and dedicated loss function (MaxLoss / MISTLoss).

[0224] In the back-learning phase, based on the borrower's actual loan application and repayment behavior data, the optimal that is presumed to have triggered such behavior β and k The value can be determined by back-calculating it. This can be formulated as an optimization problem that explores the latent variable value that best explains the observed behavior.

[0226] As an effect of this embodiment, future risks can be predicted using game theory by quantifying the borrower's intrinsic behavioral characteristics solely through objective financial logs, without the need for subjective survey data.

[0228] This embodiment can explain the final decision-making process for optimizing the financial company's profits based on the derived sensitivity.

[0230] First, in the expected return function construction step, the system, based on the individual estimated in Example 2 β and k Using as a variable, a curve combining expected interest income and expected default loss based on interest rate changes can be generated in real time.

[0232] Next, by differentiating the profit curve generated in the optimal interest rate calculation step or applying a numerical optimization algorithm, the point where the slope becomes zero—that is, the point where the financial company's profit is maximized while controlling risk—can be automatically calculated as an individually customized optimal interest rate. In this case, rather than simply raising the interest rate because the risk is high, the rate can be determined by including in the calculation the probability that the customer will churn if the rate is raised.

[0234] The optimal interest rate value calculated during the system integration and execution phases is transmitted to an external credit screening system via a control API and can be presented to customers as a real-time loan offer.

[0236] As an effect of this embodiment, the overall asset soundness and profitability of a financial institution can be simultaneously improved by offering attractive interest rates to prime customers to prevent them from leaving, and appropriate interest rates to high-risk customers to mitigate their risks.

[0238] According to some embodiments, the system of the present invention may incorporate a time series model that predicts changes in the borrower's future economic situation. By utilizing a Long Short-Term Memory (LSTM) network, the probability distribution of future income can be predicted using past income history, job stability indicators, and industry business indices as inputs. This allows for the preemptive reflection of future risk changes in long-term loans and enables dynamic interest rate adjustments according to the business cycle.

[0240] According to some embodiments, the system of the present invention can extend single loan product optimization to perform optimal interest rate determination considering the financial institution's entire loan portfolio. A portfolio risk management module that analyzes correlations between individual loans, measures concentration risk, and calculates Value at Risk may be added. Through a multi-objective optimization engine, a Pareto optimal solution can be sought that simultaneously considers profitability maximization, risk minimization, and portfolio diversity maintenance.

[0242] According to some embodiments, the system of the present invention may incorporate a reinforcement learning agent to reflect changes in market conditions and feedback on borrower behavior in real time. The state may include the current portfolio state, market interest rates, and borrower characteristics, the behavior may be the loan interest rate to be offered, and the reward may be the realized profit minus a risk penalty. The policy network may be trained using Proximal Policy Optimization or a Soft Actor-Critic algorithm.

[0244] According to some embodiments, the system of the present invention can perform distributed learning without transmitting personal information to a central server by applying a federated learning framework. By aggregating only the local model parameters of each branch or partner company and applying differential privacy, the risk of personal information leakage can be mathematically limited. This enables collaborative learning among multiple financial institutions while complying with financial regulations.

[0246] According to some embodiments, the system of the present invention can quantify the contribution of each characteristic to interest rate determination by integrating a SHAP (SHapley Additive exPlanations)-based characteristic importance analyzer. Additionally, through a counterfactual explanation generator, it can explain how interest rates change when conditions change and suggest ways to improve the borrower's interest rate. This strengthens financial consumer protection and facilitates responses to audits by regulatory authorities.

[0248] The present invention can provide an innovative system and method for determining an optimal loan interest rate for an individual by organically combining causal inference, game theory, behavioral economics, and deep learning. A three-stage pipeline that eliminates selection bias using Double ML-based causal inference, estimates unobserved economic propensity using latent parameter inversion techniques, and resolves strategic interactions using a game-theoretic equilibrium model may be a key differentiating factor of the present invention.

[0250] This invention can provide multidimensional value, such as enhancing profitability in the financial industry, advancing risk management, improving customer experience, and strengthening regulatory compliance. It can maximize expected returns through personalized interest rates, ensure asset soundness through sophisticated risk prediction based on causal effects, and increase customer trust through a fair and transparent interest rate determination process. Furthermore, explainable AI can facilitate compliance with financial regulations and respond to audits.

[0252] The system and method of the present invention can be applied to various financial products, such as personal credit loans, loans to small and medium-sized enterprises, mortgage loans, and automobile installment financing, and can be extended to pricing problems outside the financial sector, such as insurance premium calculation and subscription service pricing. Through this, the present invention can be evaluated as an innovative invention that opens new horizons for financial AI technology. Explanation of the symbols

[0254] 100: Multidimensional Financial Behavior Data Preprocessing and Adaptive Sampling Module 110: Heterogeneous Data Integrator 120: Behavior-based Intrinsic Tendency Extractor 121: Sparse Data Correction Engine 122: Full data pass filter 130: Step-by-step Adaptive Balance Sampler 200: Latent Parameter Estimation and Double ML Engine 210: 1st Learning Unit 220: 2nd Learning Unit 230: Latent Variable Inverter 300: Game Theory Equilibrium Conditions Implicit Decision Modeling Module 310: Game Theory Equilibrium Constraint Conditional 320: MaxLoss & MISTLoss Optimizer 400: API for Optimal Interest Rate Determination and Control Based on Expected Return Maximization 410: Expected Return Function Generator 420: Profit Maximizing Interest Rate Calculator 430: Real-time Loan Policy Integration Interface

Claims

Claim 1 A system for determining an optimal loan interest rate by estimating the responsiveness of individual borrowers to changes in interest rates, comprising: a data preprocessing unit that synchronizes the borrower's credit information and financial behavior data based on T1, the time of loan approval; a Double ML-based residual extraction engine that extracts respective prediction residuals from a first learning model that predicts the probability of delinquency (y) based on borrower characteristics (X) and a second learning model that predicts the interest rate (T) proposed by the financial institution; and a latent parameter estimation unit that includes a MISTnet artificial neural network structure that aligns the result residual derived from the first learning model with the intervention residual derived from the second learning model, and estimates the time discount rate and delinquency penalty sensitivity—unobserved implicit economic tendencies of the borrower not explicitly appearing in the borrower characteristics (X)—by inversion as latent variables based on the extracted residuals. A deep learning-based individual interest rate sensitivity estimation and optimal loan interest rate determination system using the same, comprising an optimization control unit that calculates an individual optimal loan interest rate that maximizes the financial company's expected profit by performing optimization operations based on individual interest rate sensitivity calculated based on the above-mentioned estimated latent variables and a game-theoretic equilibrium model containing Nash equilibrium conditions, and including an economic constraint condition in the loss function that the probability of loan execution monotonically decreasing as interest rates rise, thereby calculating an individual optimal loan interest rate that maximizes the financial company's expected profit. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 A deep learning-based individual interest rate sensitivity estimation and optimal loan interest rate determination system using the same, wherein, in claim 1, the optimization control unit uses a full dataset not used in the learning phase of the first learning model and the second learning model as the input to a residual regression model for calculating the optimal loan interest rate. Claim 6 A method for determining an optimal loan interest rate performed by the optimal loan interest rate determination system of claim 1, comprising: a step in which a data preprocessing unit collects financial behavior data of a borrower and synchronizes a time series based on the time of loan approval; a step in which a residual extraction engine applies equilibrium sampling to the synchronized data to extract the borrower's delinquency residual and interest rate treatment residual, respectively; a step in which a latent parameter estimation unit estimates the borrower's latent economic parameters and interest rate sensitivity from the extracted residuals through an artificial neural network containing game-theoretic equilibrium conditions; and a step in which an optimization control unit determines an individual optimal loan interest rate by searching for the maximum point of an expected return function with the estimated interest rate sensitivity as a variable under monotonically decreasing economic constraints included in the loss function. Claim 7 In claim 6, the step of estimating the parameters and interest rate sensitivity is characterized by integrating an economic decision-making function, in which a borrower decides whether to accept or repay a loan to maximize their utility, into the activation function of the artificial neural network, thereby providing a deep learning-based method for estimating individual interest rate sensitivity and determining an optimal loan interest rate using the same. Claim 8 A deep learning-based method for estimating individual interest rate sensitivity and determining the optimal loan interest rate using the same, further comprising the step of transmitting the determined optimal loan interest rate data to an external credit screening engine via an API (Application Programming Interface) to execute a real-time credit policy in claim 6. Claim 9 delete Claim 10 delete

Images