AI4ML for banking sector
AI4ML — Machine Learning and Analytics Platform
1. An AI Machine Learning (ML) model that can predict individual credit delinquency patterns & trends over a specified period
Using a dataset of 100,000 banking clients, we set out to demonstrate or to predict the Credit Delinquency patterns and trends across this number. We used 70,000 of the dataset as the source training data for the model and 30,000 as the test data.
Training Dataset: 70 000 instances
Test/Control Dataset: 30 000 instances for testing
Results:
The results are fascinating, the system predicted that overall 646 out of 30,000 will experience serious delinquency over the next 2 years and this was proved as accurate 392 times. This provides 57% accuracy and implies that it is accurate 57 out of 100 cases in predicting credit delinquency patterns.
2. Bank Product Recommendation engine
In this video, we tried to create ML model that is able to predict that customer will purchase the targeting product of a bank over the next 18 months. The experiments have been made on 3 major products: Payroll account, E-account, Direct Debit
Training Dataset: 58 004 instances for training
Test/Control Dataset: 24 800 instances for testing
Results
- Payroll account
Number of instances of class 0 predicted as 0 = 24157, Number of instances of class 0 predicted as 1 = 44, Number of instances of class 1 predicted as 0 = 239, Number of instances of class 1 predicted as 1 = 330
- E-account
Number of instances of class 0 predicted as 0 = 23128, Number of instances of class 0 predicted as 1 = 95, Number of instances of class 1 predicted as 0 = 130, Number of instances of class 1 predicted as 1 = 1506
- Direct Debit
Number of instances of class 0 predicted as 0 = 23975, Number of instances of class 0 predicted as 1 = 44, Number of instances of class 1 predicted as 0 = 768, Number of instances of class 1 predicted as 1 = 73
3. Predicting final loan status of a creditor
This example is similar to what we have done with Credit Delinquency data. The goal is to classify whether the creditor will fully pay the loan in time or will give up on trying to collect unpaid debt.
Traning dataset: 70 000 instances
Test/Control Dataset: 30 000 instances
Results
Number of instances of class “Charged Off” predicted as “Charged Off” = 1619, Number of instances of class “Charged Off” predicted as “Fully paid” = 5143, Number of instances of class “Fully paid” predicted as “Charged Off” = 268, Number of instances of class “Fully paid” predicted as “Fully paid” = 22976
The model provides 81% accuracy and implies that it is accurate 1618 out of 1881 cases in predicting Failure to pay a loan patterns
4. Predicting Housing prices
Based on a 3-year history of housing related, we attempted to build a model that accurately predicts housing prices.
Training Dataset: 21 329 instances
Test/Control Dataset: 9 142 instances
Results:
The average difference between predicting price and actual price was 1 384 483. This may be considered as a good result since prices vary from 5 million to 20 million
AI4ML — Machine Learning and Analytics platform. www.ai4ml.com
