2. Some interesting facts
► Spikes in the churn rate at the end of the deposit period
► We have to adapt the active time line for every customer
► The last 3 months before the churn are the most
informative
► Different account types have different patterns in churn
► Slow attriters pay down their outstanding balance until
they become inactive
► Fast attriters quickly pay down their balance and either
lapse it or close it
► Proper segmentation of the customers is the key for
further precise modeling
3. Groups of variables
► Account transaction IN
► Account transaction OUT
► Service indicators
► Personal profile
► Customer level information
4. Basic attributes
► Age
► Sex
► Education
► Income
► Occupation
► Service starts
► Asset to liability ratio
► Date of account opening
5. Business variables
► Deposit accounts
► Deposit balance
► The number of deposits
► Monthly amount of consumption
► Contract ending date
► Decreasing account balance
► Decreasing number of credit card
purchase
► Customer profitability for the bank
base on duration and age
► Number of phone payments
► Number of ATM transactions
► Number of card payments
► Time of last transaction
► First contact date
► Average debit amount per month
► Average credit amount per month
► Status: active, not active, churn
► Account type (segment): saving,
current, loan, deposit etc.
► Spending speed
► Annual charge date
► Maximum transaction amount
► Salary account
6. Models’ review
Approach
Best
model Article
CART + TreeNet +
C5.0
CART
Chandar, M Laha, A., and Krishna, P. Modeling churn behavior of bank
customers using predictive data mining techniques. [J] National Conference on
Soft Computing Techniques for Engineering Applications (SCT-2006). 2006, (3)
24-26
FUZZY C-Means
clustering
N/A
Popovic,D. and Basic B.D. Churn prediction Model in retail banking Using Fuzzy
C-means Algorithms .[J] Informatica 2009 (33) 243-247
Sampling techniques
+ gradient Boosting +
weighted random
forest
Weighted
random
forest
Buzes, J. and Poel D.V. Handling Class Imbalance in Customer Churn
Prediction[J] Expert System with Applications 2009 (36) 5445-5449
SVM + random
sampling +
imbalance
characteristic of data
set
N/A
Benan He, Yong Shi, Qian Wan, Xi Zhao, Prediction of customer attrition of
commercial banks based on SVM model. Procedia Computer Science 31 (2014)
423-430
7. Models’ review
Approach Good performance Article
Artificial –based classification
(ABC) using Artificial Natural
Networks (ANN)
3
Fan Li, Juan Lei, Ying Tian, Sakuna
Punyapatthanakul, Yanbo J. Wang
Model Selection Strategy for
Customer Attrition. Risk Prediction in
Retail Banking. Proceedings of the 9th
Australian Data Mining Conference.
Bayesian-based classification
(BBC) or Naïve Bayes (NB)
6
Case-Based classification
(CBC) k-NN
4
Rule-based classification
(RUBC)
1
Regression based
classification. (RBC) logistic
regression (LR)
5
Tree-based classification
(TBC) –greedy algorithm
2