運用機器學習方法推廣綜合券商大財管業務

EXPAND INTEGRATED SECURITIES FIRM'S BUSINESS USING MACHINE LEARNING

10.6338/JDA.202206_17(2).0002

許仲廷(Zhong Ting Xu)

證券 ； 客戶分群 ； 分類 ； 機器學習 ； k mean ； logistic regression ； Stock Repurchase ； Capital reduction by cash ； Event Study ； Multiple Regression Analysis

Journal of Data Analysis

17卷2期（2022 / 06 / 01）

31 - 58

繁體中文

本研究希冀以統計以及機器學習之方法找出統一證券財管三大業務（保代、海外市場、財富管理）之客戶特徵，及區分有無購買客群間之關鍵差異性質，使業務員得以用更有效率、有系統之方式辨別新開戶商品偏好，藉此提高開戶、成交而後達到提高營收之成效。研究流程從上述財管三大類業務的k-means分群開始，藉此完整建立大財管客戶之客戶素描以及兩年間（2018、2019）之特性轉變研究；第二步將未曾購買大財管商品之證券客戶與大財管客戶混和，使用分類演算法訓練模型，並探討模型中統計顯著之變項對客戶購買行為之影響，讓模型面對新開戶之特性能夠分辨其購買大財管商品之概率。在羅吉斯迴歸分類中，兩組模型顯著的變數相同，都是年齡、性別、註冊天數、開戶在2008前以及開戶在年終發放月份這五個變數，而購買總金額意外的並不顯著；從分類結論可以篩除潛在購買可能低的無效用戶，專注於潛在客戶，並依分群結論優化不同客群之服務，提高成交、營收。

This study is designed to use statistic and machine learning methods to find out the customer characteristics of President securities' three major financial management businesses (Insurance, Overseas, Wealth management), and to distinguish the key differences between the customer groups, so that salesperson can use this to identify the preferences of new customers in a more efficient and systematic way, thereby improving new comer rate and transaction rate, then achieving the goal of increasing revenue. The process starts with the k-means clustering of the three financial management business' customers, so as to establish the customer sketches of these businesses and research on the change in characteristics of customers during data time (2018, 2019). Second step, use Logistic classification to train the model in order to distinguish from buyers(1) to non-buyers(0), and discuss how the statistically significant features in the model impact on customer buying behavior, so that the model can tell the new comer buying probability in the future. In the outcome of Logistic classification, the significant variables of the two models are the same, which are Age, Gender, Days_of_registration, Registered_before_2008, Registered_at_bonus_release. And Total_buy_amount is interestingly not significant. In conclusion, it is possible to screen out users with low potential purchasing probability by focusing on potential customers, and optimize the services of different customer groups.

1. (2011)。我國證券商之發展與當前監理方向。證券暨期貨月刊，29(9)
2. Stack Exchange, (2016), User segmentation by clustering with sparse data, https://stats.stackexchange.com/questions/199501/user-segmentation-byclustering-with-sparse-data
3. Big Data Science Practice, (2015), Impact of target class proportions on accuracy of classification,https://kumaranpm.blogspot.com/2015/03/impact-of-target-class-proportions-on.html
4. Wikipedia - Statistical classification , https://en.wikipedia.org/wiki/Statistical_classification
5. Building A Logistic Regression in Python, Step by Step,https://towardsdatascience.com/building-a-logistic-regression-in-pythonstep-by-step-becd4d56c9c8
6. Wikipedia - Market segmentation, https://en.wikipedia.org/wiki/Market_segmentation
7. Wikipedia - k-means clustering, https://en.wikipedia.org/wiki/Kmeans_clustering
8. https://www.geeksforgeeks.org/ml-logistic-regression-using-python/
9. Huang, D., (2017)，你可能不知道的邏輯迴歸(Logistic Regression)，https://taweihuang.hpd.io/2017/12/22/logreg101/
10. Huang, J., (2019) ， k means 分群演算法與 Silhouette 輪廓分析，https://ppt.cc/f8lhyx
11. Jordan, J., (2017), Customer Segmentation Based on Purchasing Behavior, https://github.com/jeremyjordan/customersegmentation/blob/master/customer_segments.ipynb.
12. Li, S.(2017).Li, S., (2017), Building A Logistic Regression in Python, Step by Step..
13. Mohit Gupta_OMG(2020).Mohit Gupta_OMG, 2020, ML | Logistic Regression using Python.
14. Prasad, G., (2019), Notes on K-prototype for clustering mixed typed data, https://medium.com/@guruprasad0o_o0/notes-on-k-prototype-for-clusteringmixed-typed-data-e80eb526b226
15. 袁道寬(2017)。國立臺北大學統研所。
16. 陳信宏(2004)。國立臺灣大學商學研究所。
17. 陸昀泓(2014)。國立臺灣大學國際企業學研究所。
18. 謝育儒(2016)。國立高雄應用科技大學金融系金融資訊碩士在職專班。