自動信用風險評分模型之研究

2008年發生百年來史無前例的金融海嘯，直接地影響全球金融體系崩潰，迫使金融業必須正視風險管理的挑戰，促使信用評分機制的重要性提高。而信用評等模型可以幫助銀行針對各式各樣不同的客戶行為，個別預測該客戶風險程度的高低。以實務經驗來探討，若銀行以傳統方式製作信用評分模型，一般需約兩周的時程，而本論文所提出之自動化的過程有別以往，可將整體時程縮減至20分鐘就可以產出評分模型，但在加速的過程中卻不會失去以往人工的準確性，自動地找出最適合的模型，使銀行內的分析師能夠更加有效率的進行分析。本研究以傳統評分卡製作標準流程，將過程分為五個部分，首先將客戶資料進行資料前處理，再將前處理好的資料進行單因子分析，接著建立初步的信用評分模型，模型建立完成後產生最終信用評分卡，產生出的評分卡將進行評估，以確保最終評分卡的準確度。本研究以Python進行後台處理和計算最優模型，再將資料存放於MySQL，最終使用PHP做前端Web介面，以最直觀的方式呈現給使用者。本研究可自動建置業界標準的信用評分模型，並可瞭解模型細節的資訊，同時可進一步對未來數據做預測。實驗結果顯示，本研究使用自動貝葉斯優化調參，可進一步提升模型預測在AUC、Gini及KS上的效能。

關鍵字

自動機器學習；評分卡；資料前處理；模型參數優化

並列摘要

The financial tsunami in 2008 directly affected the collapse of the global financial system, forced the financial industry to face up to the challenges of risk management, and promoted the importance of credit scoring mechanisms. The credit rating model can help banks to individually predict the risk level of a customer for a variety of different customer behaviors. Based on practical experience, if a bank uses the traditional way to make a credit scoring model, it usually takes about two weeks, and the automated process proposed in this paper is different from the past, and the overall time can be reduced to 20 minutes. The scoring model is developed, but the previous manual accuracy will not be lost in the process of acceleration, and the most suitable model will be automatically found, so that analysts in the bank can conduct analysis more efficiently. This research uses the standard process of traditional scorecard production and divides the process into five parts. First, the customer data is pre-processed, and then the pre-processed data is subjected to single-factor analysis, and then a preliminary credit scoring model is established, and the model is completed. After the final credit scorecard is generated, the generated scorecard will be evaluated to ensure the accuracy of the final scorecard. This research uses Python for background processing and calculation of the optimal model, then stores the data in MySQL, and finally uses PHP as the front-end web interface to present it to the user in the most intuitive way. This study can automatically build an industry-standard credit scoring model, understand the details of the model, and make further predictions about future data. The experimental results show that the use of Bayesian optimization parameters in this study can further improve the performance of model prediction on AUC, Gini and KS.

並列關鍵字

Automated Machine Learning ； Scorecard ； Data Preprocessing ； Model Parameter Optimization

主題瀏覽