Many financial institutions have tried to develop better risk assessment models for unsecured personal loans. However, traditional methods usually do not consider the decision-making process of applicants who select and execute loans based on specific loan terms (e.g., interest rates and limits). This can lead to misaligned risk assessments and, eventually, bank losses. This paper introduces an end-to-end multi-task learning framework that simultaneously predicts default and loan selection probabilities, focusing on fine-tuning risk assessments for applicants who select loans. We adjust the loss of default prediction for applicants with a high selection probability and apply linear modulation to better integrate credit data and loan terms for loan selection prediction. This approach enhances the influence of loan terms relative to other features, allowing the model to better capture how these terms impact applicant decisions. We demonstrate that our method outperforms other machine learning and deep learning approaches in predicting defaults for applicants who select loans, while also showing consistent performance stability across different months. To validate our framework, we use large-scale, real-world loan data from Shinhan Card, one of the largest credit card companies in Korea.