효용함수 추정을 통한 최적 신용카드 오퍼 모형의 개발

AAs credit card industry in Korea is expanding and competition among credit card companies grows up, reaction rate of mass marketing quickly decreases and rate of cost of advertising campaign to customer increases. Many companies are interested in decreasing the cost and try to build reaction ‐prediction models. They often use logit, probit as parametric estimation model, and neural network, scorecard model as non‐parametric estimation model. In this paper, as we set up the individual utility as function of credit card characteristics, total limit, IS interest rate, CS interest rate, marketing offer vector and customers’ profile vector, we are finding the optimal offer of credit card. To find optimal model, we compare several models, logit, probit as parametric model and neural network, scorecard model as noon‐parametric model. Furthermore, to find optimal offer to customer, we calculate marginal rate of substitution between two offers which are usually used. Neural network model use linear function as combination function, hyperbolic tangent as activation function, three hidden layers and back‐propagation network as error correction. Estimation result, directions of coefficients of five models are alike and important input variables are marketing offer, using amount of IS, male’s total limit, IS using customer’s CS and IS interest rate. In training sample, neural network model is the best at MAD, RMSE, Log‐Likelihood, Rate of Correct Prediction and Cost‐Benefit Analysis. But in validation sample, parametric models (logit, probit) are better than others. Because of neural network model’s over‐fitting problem, it’s performance decreases more in validated sample than in training sample compared to other models. In terms of MRS, customers with same profile have same MRS in parametric models, but different MRS in neural network model. We found weak evidences that neural network model is most efficient in terms of calculating MRS. To improve this study, we need accumulation of panel data of marketing activities; the same marketing performed to the same customer several times with a lag. Furthermore, the real characteristics of credit card inside of customer’s wallet is available, the prediction rate of the model would be improved. Survey data would be good candidate.