Dominik Matula 23th June, 2020
Instalment Detector
Revealing clients’ behaviour from transactional data
Instalment Detector
Revealing clients’ behaviour from transactional data
sli.do
#PROFINIT
Dominik Matula
Senior Data Scientist
Specialization: Finance, Banking
Research: Pseudosocial networks
VAT Fraud detection
Use case
4
Your bank
Use-case
Wild
world
Greta Karl Mark
LOAN LOAN LOAN
LOAN
LOAN
LOAN
LOAN
LOANLOAN
LOAN
LOAN LOAN
5
Key goals of
Instalment detector
ADVANCED
RISK SCORING
CLIENT‘S
LOYALTY
PROFIT
MAXIMIZATION
€€€
CLIENT‘S
SAVINGS
€€€
Information mining
7
How to reveal an instalment transaction?
Client Receiver Date € Symbol Note
Karl 123456 2020-04-18 21.45 Loan payment
Anna 513487 2020-05-15 172.99 4783351
Greta 363391 2020-03-21 45.87 76558 Fancy shoes
Thomas 513487 2020-03-14 3000.00 Paying the rest
Receiver
black/white lists
Transaction
note mining
Symbol-based
triggers
› Conventional approaches
OUTDATING
CONFUSING
SPARSE
CONFUSING
SPARSE
Typical amount
filtering
IMPRECISE
8
What if …
› …a transaction doesn‘t contain enough information?
› Use complete RELATIONS
+ Information from all transactions
+ Time series patterns
+ Variability (amounts, symbols)
+ …
Client Receiver Date € Symbol Note
Karl B (new) 2020-06-18 21.45 6612162909 Mobile
Karl B (new) 2020-05-16 21.45 6612162909 Loan payment
Karl B (new) 2020-04-17 21.45 6612162909 Iphone
Advanced feature
engineering
Machine learning
10
Results
Battle Plan
Super-duper
features
+ power
- - interpretability
- power
+ interpretability
Black box
Oldies but
goldies
Bayesian
networks
11
Bayesian models intro
Super-duper
features
Id Feature Meaning Adjustment
F1 Amt Time Series Pattern 13 : 2
F2 Specific symbol presence 3 : 2
F3 Note contains loan 12 : 1
F4 Some other cool feature 8 : 1
… … …
PRIOR ODDS: 1:49
(Prior probability: 2 %)
Got those from
TRAINING data
# CASES w F
# nonCASES w F
12
Prior probability
F2 F4
From Naive Bayes to Bayesian network
F1 F3 F5
Posterior probability
⋅
3
1
⋅
3
2
⋅
12
1
⋅
13
2
Id Meaning Adj.
F1 Young sender 3 : 1
F2 Amt Time Series Pattern 13 : 2
F3 Constant Symbol 3 : 2
F4 Note contains 'loan' 12 : 1
F5 Some other cool feature 8 : 1
1
49
⋅
8
1
2 %
86 %
Let‘s get advanced
14
› .. is a multilayer Bayesian network model
– RELATION layer
– RECEIVER layer
– Other layers if needed: SENDER, COUNTER-RELATION, …
Instalment detector
RELATION layerRECEIVER layer
15
Other context examples
› RECIEVER context based knowledge
› COUNTER-RELATION context based knowledge
Client Receiver Date € Symbol Note
Karl 123456 2020-04-18 21.45 32154342 Loan payment
Anna 123456 2020-05-15 73.50 99554722
Greta 123456 2020-03-21 45.87 32154688 Loan TV
Sender Receiver Date € Symbol Note
123456 Karl 2020-04-15 5000.00 4783325 cons loan 2020/04
Karl 123456 2020-05-15 43.20 4783325
Karl 123456 2020-06-15 43.20 4783325
Achieved results
17
› Implemented in a major European bank (2018)
– Billions of transactions, ca. 1 TB of data
– Hive+Spark, modelled in R
– Further implementations discussed in several other banks
(Germany, Czechia)
› Intended boost: +10%
› Reality: +100%
– Small lending companies
– Changes in banking accounts
– P2P loans (family, friends, …)
– Indirect instalments
– …
Results
1.56M
17k
1.48M
BANK Intersection Extra
INSTALMENT DETECTOR
NumberofInstalmentPayments
271k
Conflict
Feature description DWH Hive Spark *)
No. of unique clients per RECEIVER 2280 s 242 s 71 s
Regex spl[aá]tk **) matching 4260 s 38 s 40 s
*) + initial overhead
**) Real feature contains more complicated expressions. Unfortunatelly, DWH couldn‘t make it…
19
Key advantages of Instalment Detector
Straightforward
INTERPRETABILITY
ADAPTABILITY
to market changes
Client Receiver Date € Symbol Note
Karl B (new) 2020-06-18 21.45 6612162909 Mobile
Karl B (new) 2020-05-16 21.45 6612162909 Loan payment
Karl B (new) 2020-04-17 21.45 6612162909 Iphone
ADVANCED
feature engineering
Directly aplicable in
OTHER FIELDS
Salary
Alimony Leasing
InsuranceGaming
Mortgages
Profinit EU, s.r.o.
Tychonova 2, 160 00 Prague 6 | Phone + 420 224 316 016
Web
www.profinit.eu
LinkedIn
linkedin.com/company/profinit
Twitter
twitter.com/Profinit_EU
Facebook
facebook.com/Profinit.EU
Youtube
Profinit EU
Thank you
for your attention!

Profinit webinar: Instalment Detector

  • 1.
    Dominik Matula 23thJune, 2020 Instalment Detector Revealing clients’ behaviour from transactional data
  • 2.
    Instalment Detector Revealing clients’behaviour from transactional data sli.do #PROFINIT Dominik Matula Senior Data Scientist Specialization: Finance, Banking Research: Pseudosocial networks VAT Fraud detection
  • 3.
  • 4.
    4 Your bank Use-case Wild world Greta KarlMark LOAN LOAN LOAN LOAN LOAN LOAN LOAN LOANLOAN LOAN LOAN LOAN
  • 5.
    5 Key goals of Instalmentdetector ADVANCED RISK SCORING CLIENT‘S LOYALTY PROFIT MAXIMIZATION €€€ CLIENT‘S SAVINGS €€€
  • 6.
  • 7.
    7 How to revealan instalment transaction? Client Receiver Date € Symbol Note Karl 123456 2020-04-18 21.45 Loan payment Anna 513487 2020-05-15 172.99 4783351 Greta 363391 2020-03-21 45.87 76558 Fancy shoes Thomas 513487 2020-03-14 3000.00 Paying the rest Receiver black/white lists Transaction note mining Symbol-based triggers › Conventional approaches OUTDATING CONFUSING SPARSE CONFUSING SPARSE Typical amount filtering IMPRECISE
  • 8.
    8 What if … ›…a transaction doesn‘t contain enough information? › Use complete RELATIONS + Information from all transactions + Time series patterns + Variability (amounts, symbols) + … Client Receiver Date € Symbol Note Karl B (new) 2020-06-18 21.45 6612162909 Mobile Karl B (new) 2020-05-16 21.45 6612162909 Loan payment Karl B (new) 2020-04-17 21.45 6612162909 Iphone Advanced feature engineering
  • 9.
  • 10.
    10 Results Battle Plan Super-duper features + power -- interpretability - power + interpretability Black box Oldies but goldies Bayesian networks
  • 11.
    11 Bayesian models intro Super-duper features IdFeature Meaning Adjustment F1 Amt Time Series Pattern 13 : 2 F2 Specific symbol presence 3 : 2 F3 Note contains loan 12 : 1 F4 Some other cool feature 8 : 1 … … … PRIOR ODDS: 1:49 (Prior probability: 2 %) Got those from TRAINING data # CASES w F # nonCASES w F
  • 12.
    12 Prior probability F2 F4 FromNaive Bayes to Bayesian network F1 F3 F5 Posterior probability ⋅ 3 1 ⋅ 3 2 ⋅ 12 1 ⋅ 13 2 Id Meaning Adj. F1 Young sender 3 : 1 F2 Amt Time Series Pattern 13 : 2 F3 Constant Symbol 3 : 2 F4 Note contains 'loan' 12 : 1 F5 Some other cool feature 8 : 1 1 49 ⋅ 8 1 2 % 86 %
  • 13.
  • 14.
    14 › .. isa multilayer Bayesian network model – RELATION layer – RECEIVER layer – Other layers if needed: SENDER, COUNTER-RELATION, … Instalment detector RELATION layerRECEIVER layer
  • 15.
    15 Other context examples ›RECIEVER context based knowledge › COUNTER-RELATION context based knowledge Client Receiver Date € Symbol Note Karl 123456 2020-04-18 21.45 32154342 Loan payment Anna 123456 2020-05-15 73.50 99554722 Greta 123456 2020-03-21 45.87 32154688 Loan TV Sender Receiver Date € Symbol Note 123456 Karl 2020-04-15 5000.00 4783325 cons loan 2020/04 Karl 123456 2020-05-15 43.20 4783325 Karl 123456 2020-06-15 43.20 4783325
  • 16.
  • 17.
    17 › Implemented ina major European bank (2018) – Billions of transactions, ca. 1 TB of data – Hive+Spark, modelled in R – Further implementations discussed in several other banks (Germany, Czechia) › Intended boost: +10% › Reality: +100% – Small lending companies – Changes in banking accounts – P2P loans (family, friends, …) – Indirect instalments – … Results 1.56M 17k 1.48M BANK Intersection Extra INSTALMENT DETECTOR NumberofInstalmentPayments 271k Conflict Feature description DWH Hive Spark *) No. of unique clients per RECEIVER 2280 s 242 s 71 s Regex spl[aá]tk **) matching 4260 s 38 s 40 s *) + initial overhead **) Real feature contains more complicated expressions. Unfortunatelly, DWH couldn‘t make it…
  • 18.
    19 Key advantages ofInstalment Detector Straightforward INTERPRETABILITY ADAPTABILITY to market changes Client Receiver Date € Symbol Note Karl B (new) 2020-06-18 21.45 6612162909 Mobile Karl B (new) 2020-05-16 21.45 6612162909 Loan payment Karl B (new) 2020-04-17 21.45 6612162909 Iphone ADVANCED feature engineering Directly aplicable in OTHER FIELDS Salary Alimony Leasing InsuranceGaming Mortgages
  • 19.
    Profinit EU, s.r.o. Tychonova2, 160 00 Prague 6 | Phone + 420 224 316 016 Web www.profinit.eu LinkedIn linkedin.com/company/profinit Twitter twitter.com/Profinit_EU Facebook facebook.com/Profinit.EU Youtube Profinit EU Thank you for your attention!