Detection of excise fraud in trading networks Sébastien Beaudoint – Lore Cloots
Excise Fraudsters Gotta Catch 'Em All #SFBL14 #Excise
SAS Forum 09 O...
Detection of excise fraud in trading networks Sébastien Beaudoint – Lore Cloots
Excise Fraudsters Gotta Catch 'Em All #SFBL14 #Excise
SAS Forum 09 October 2014
Outline Introduction Detection of excise fraud in trading networks
Outline Introduction Detection of excise fraud in trading networks
Introduction
Federal Public Service Finance: 23.409 people Customs & Excise: 3.659 people Investigation & Research: 300 people
Dataminers: 2 people
Outline Introduction Detection of excise fraud in trading networks
Problem definition Excise products Decided at European Level Alcohol (Beer, Wine, Sparkling wine, Spirituous, Intermediate products) Energy products (Petrol, Electricity) Tobacco Minimal rate decided at European Level
Final rate decided by every member state
National excise products Decided by every member state For Belgium Non-alcoholic beverages (except milk) Coffee
Problem definition Rate difference per member state Beer
Wine
Problem definition EU legislation Movements under suspension of excise duty (= duties not yet paid)
Problem definition Evasion at the end of the duty-suspension No payment
Payment in a wrong country (“Excise low level”)
Problem definition Evasion at the end of the duty-suspension Smuggling No document
Multiple use of one document
Problem definition EMCS (Excise Movement Control System) E-AD (Electronic administrative document)
Problem definition: movement overview
Problem definition: project Observation: Smugglers are interconnected Use of social network analysis concepts in the project
Problem definition: project Approach: Risk score Specification of each Belgian company working in the system Number of shipments Number of shipments to low/medium/high excise rate level country Status of shipments Stock
Social Network variables Number of shipments with suspicious company Centrality, betweenness …
E-miner model Decision tree Neural network …
Outline Introduction Detection of excise fraud in trading networks
Data preparation EMCS (Excise Movement & Control System) Information about movements of excise goods under suspension of excise: Excise Product Consignor Consignee Quantity Alcoholic strength Degree Plato (°P, for beer) …
AC4 Information about payment of excise products Excise Product Declarant Quantity in tax base value (depending on excise products: Hl, HlAlc, Hl°P)
Data preparation: Consolidation One line per company ± 2.600 consolidated variables Authorization information Global consolidation Last year of activity consolidation Absolute and relative variables …
Data preparation: problems encountered Data evolution in time Application evolved so the data available also evolved
Missing data & data coherence Solutions Including evolution information in the consolidation Trying to reconstruct missing or incoherent information Including variables indicating data problems (might be a fraud indicator)
Outline Introduction Detection of excise fraud in trading networks
Data exploration In reality, many cycles of data preparation - data exploration Help to check potential use of calculated variables Help to understand some input data
Variable selection Best variables to determine the target variable
Outline Introduction Detection of excise fraud in trading networks
Divide input data into training and validation samples Define specific models and parameters Compare results
Modeling: models comparison ROC curve
Modeling: problems encountered Quantity & Quality of target data Few positive fraud historical info (± 1 %) No fraud status is poor assumption One model but different fraud behaviours
Solutions Oversampling, under sampling, SMOTE (Synthetic Minority Over-sampling Technique) Probability Matrix and Cost Matrix Creation of different models Going back to fraud history to gather more information …
Outline Introduction Detection of excise fraud in trading networks
Deployment: web application
Deployment: web application
Deployment: web application
Deployment: web application
Deployment: web application
Deployment: web application
Deployment: web application
Deployment: web application
Deployment: web application
Questions?
Thank you for your attention
Questions?
SAS Forum
Twitter Contest – Tweet to win prizes!
3. In which country do you pay the least excise duties for beer?
Belgium
France
Ireland
Prizes to win:
Tweet your answer:
1st prize: a ticket for Analytics 2015 2nd prize: a book of Prof Bart Baesens: “Analytics in a big data world” 3rd to 30th prize: chocolates with pepper