Our team chose the project from Credit Suisse, and developed a machine learning tool identifying clients that are likely to be involved in money laundry activities and to tell them apart from good clients.
- clean_code: notebooks with the final code producing important results
- notebooks: exploritary notebooks which are not included in final result
- laundrywomen_presentation: our pitch presentation
===
+++ DATA SOURCES
Data provided by CreditSuisse:
'large.csv' (http://bit.ly/36jCXYi)
'small.csv' (http://bit.ly/33kJboM)
'country.csv' (http://bit.ly/36nQ5f4)
'jeopardy.csv' (http://bit.ly/2NCgFZ8)
Open source data:
- Basel AML Index: independent annual ranking that assesses the risk of money laundering and terrorist financing around the world https://www.baselgovernance.org/basel-aml-index/public-ranking
'AML_risk_ranking.csv'
- Corruption Perceptions Index 2018 by Transparency International: ranks 180 countries by their perceived levels of public sector corruption according to experts and businesspeople https://www.transparency.org/cpi2018
'corruption_index_cross-over.csv'
+++ DATA PRE-PROCESSING
-
Feature augmentation log and norm transformations
-
Introduced 4 categories: 0: Normal individual 1: Company 2: Other institution 3: is_pep individual
-
Adding new features from external sources
0: AML risk score and rank for countries
1: Corruption index for countries
+++ DATA EXPLORATION
general statistics
- type of data
- structure and relationship of features
visualize data
test outlier detection method LOF
+++ MODELS
Model choice: to understand the model's behaviour
DecisionTreeClassifier CatBoost SVM
Steps:
- train-test split
- upsampling of minority class in training set
- cross-validation on training set; aim: optimize f1_micro score (detect minority class)
- test with test set
- evaluation: classification report, ROC
+++ INTERPRETATION
- Visualization feature importance from CatBoost and DecisionTree
- Plotting distribution of most important features
- 2D-distribution of highly interaction features
- Confirming / rejecting hypothesis about external data sources (+ CIP, - AML)
- Usual outlier detection method (LOF) is not suitable to detect suspicious behaviour