Propensity Modelling for Banks
0 likes1,304 views
This document is a presentation by Lukáš Dvořák discussing data-driven lending and propensity modeling for banks, emphasizing the shift from traditional banking to data-centric approaches. Key highlights include understanding the value of data in banking, effective client targeting using behavioral models, and monetizing data through improved propensity-to-buy scoring. The presentation also touches on various data sources, customer segmentation, and the application of advanced models to enhance lending strategies.
![41
Case Study - SUMMARY
TARGET
rate
CNT_CL_EVER
0 1 [2; 3] [4; 7] [8; 15] >= 16
CLN_SEG
ACT 4,8% 24,8% 39,2% 54,6% 73,9% 92,7%
SAC 2,6% 18,9% 31,0% 44,9% 63,3%
NEW 5,5% 27,8% 57,9%
NAC 1,2% 9,1% 17,7% 27,4% 41,1%
SAV 0,3% 3,2% 9,8%
CLOSED 0,8% 2,7% 4,7% 8,6%
% CLN Target rate AUC
31% 1,2% 0,75
1% 70,9% 0,59
68% 12,4% 0,84
Inactive & Savers
AUC .75
Frequent borrowers
AUC .59
Active & Newcomers
AUC .84
31 % of clients 68 % of clients 1 % of clients](https://image.slidesharecdn.com/webinarlending2021v06-210323112237/85/Propensity-Modelling-for-Banks-41-320.jpg)
Propensity Modelling for Banks
- 1. Lukáš Dvořák March 2021 Data-driven lending Propensity modelling for banking Online Webinar
- 2. 2 Lukas Dvorak Business Development Manager, PROFINIT HELPING CLIENTS TO BUILD BIG DATA AND DATA SCIENCE COMPETENCY
- 3. 3 Improving client targeting with a propensity model TRADITIONAL MODEL SAVERS & LESS ACTIVE CUSTOMERS MAINSTREAM CLIENTS + NEWCOMERS REGULAR BORROWERS IMPROVED MODEL
- 4. 4 What else will you learn? 1. What bank 4.0 is? 2. What data is the most useful 3. How to approach data-driven lending 4. How to monetize your data
- 6. 6 „Banking everywhere, never at a Bank.“ Brett King
- 7. 7 Ten years ago, people went to the bank to get cash. Nowadays, we don’t go to banks at all. We just send our data. Banks are not as before. They are IT companies now.
- 8. 8 Isn’t the data in a bank’s warehouse more valuable than the deposits in its accounts these days?
- 9. 9 Relationship between banks and clients is changing DATA DATA SCIENCE BIG DATA Data competency is key to clients satisfaction BANK
- 11. 11 „In our case we cannot do it, because“ „We need to focus on core business now.“ „We are a special case. AI doesn’t work for us; we trust in our intuition.“ „Data use isn’t allowed here.“ YES WE CAN
- 12. THE DATA
- 13. 13 Know Your Customer …like you know your partner
- 14. 14 Next-level data modelling BEHAVIORAL MODEL TRAINED USING TRANSACTIONAL HISTORY CLIENTS METRICS STATISTICAL FEATURES TRADITIONAL MODEL TRAINED USING CLIENTS METRICS SCORE
- 15. 15 Transactional data – ATMs › ATMs withdrawals – The most boring data in the bank? › Let’s try to explore the time series – Average number of withdrawal per hour in a week Avg number of ATM withdrawal per hour City Subway Males Other Other Station Subway
- 16. 16 Avg number of ATM withdrawal per hour Transactional data – ATMs: Factory areas
- 17. 17 Avg number of ATM withdrawal per hour Transactional data – ATMs: Business centers
- 18. 18 Avg number of ATM withdrawal per hour Transactional data – ATMs: Nightlife
- 19. 19 Pseudo-social network Connecting people with same behaviour ATMs Salary Cards Online Shopping Money Transactions Payment Abroad
- 20. 20 Well, this is pretty much the same as the real social context… Pseudo-social network Connecting people with same behaviour .73 .21 .98 .54
- 21. 21 Know your customer …like you know your partner TRANSACTIONS & CLIENT DATA SIMILARITIES MICROSEGMENTS OF „ONE“ › By getting a holistic customer view
- 22. 22 P2L › Client data: age, region, education, risk scores, etc. › Product history: loans, investments, insurance, etc. › Account data: monthly aggregates, balances, etc. › Transactional data: card payments, in/out transactions › Microsegmentation – Spending behaviour – Mobility – Use of services – Salary and incomes – Family, households and friends Important Data Inputs
- 27. 27 P2L INSTALMENTS DETECTION SOCIAL GROUPS SALARY SPENDING HOUSEHOLD BEHAVIOURAL MODEL INPUT TRANSACTIONAL DATA PROPENSITY-TO-BUY SCORING LIFE-EVENTS MONITORING HOUSEHOLD EMPLOYMENT SPENDING OPPORTUNITIES TO OFFER EVENT TRIGGERS LOAN REFINANCING DATA-DRIVEN CAMPAIGNS INPUT CLIENT DATA TARGETED LOAN OFFERS Propensity-to-buy modelling
- 28. 28 P2L Automated scoring Outcomes Propensity-to-loan score for each client individually Probability that a client takes a consumer loan Selection and ranking of clients for sales campaigns › Input data – Client data, products, monthly aggregates – Transactional data, pseudo-social networks – Microsegmentation Shopping behaviour, mobility, billing for services, salary and incomes Propensity-to-buy modelling
- 29. 29 Data Science platform in a company environment DATA SOURCES BIG DATA PLATFORM DWH STREAMING MLOps WORKFLOW MNG. DevOps ENGINEER CI/CD VERSIONING SYSTEM RETAIL DATAMART BI ANALYST CAMPAIGN MNG. TOOLS BUSINESS USERS MODEL MNG. TARGET POOL RECOMMEDER TOOL CALLSCRIPT GENERATION MODEL REPOSITORY DATA SCIENTIST SCORING
- 30. MONETIZING DATA
- 31. 31 Propensity-to-buy modelling Model as a campaign selection tool 1. Effective use of a thin channel 2. Selection of responsive clients 3. Next-best offer % 1 : 20 1 : 5 1 : 40
- 32. 32 TRADITIONAL MODEL SAVERS & LESS ACTIVE CUSTOMERS MAINSTREAM CLIENTS + NEW COMERS REGULAR BORROWERS Propensity-to-buy modelling BASELINE THE VALUE INCREASE WITH AN IMPROVED TARGETING
- 33. 33 IMPROVED MODEL TRADITIONAL MODEL SAVERS & LESS ACTIVE CUSTOMERS MAINSTREAM CLIENTS + NEW COMERS REGULAR BORROWERS Propensity-to-buy modelling BASELINE THE VALUE INCREASE WITH AN IMPROVED TARGETING
- 34. 34 › 𝑿 – Channel capacity › 𝑹𝟎 – Baseline response rate › 𝑳 – Lift of response rate › 𝒊𝒏𝒄 – Unit income Propensity-to-buy monetization example
- 35. 35 1 5 10 0% 10% 20% 30% 40% 50% Velikost výběru Lift Model 1 Model 2 ▼ › 𝑿 – Channel capacity › 𝑹𝟎 – Baseline response rate › 𝑳 – Lift of response rate › 𝒊𝒏𝒄 – Unit income Propensity-to-buy monetization example baseline Sample of the whole client base
- 36. 36 1 5 10 0% 10% 20% 30% 40% 50% Velikost výběru Lift Model 1 Model 2 ▼ Value: › 𝑿 – Channel capacity › 𝑹𝟎 – Baseline response rate › 𝑳 – Lift of response rate › 𝒊𝒏𝒄 – Unit income 𝑿 ∙ ∆𝑳 ∙ 𝑹𝟎 ∙ 𝒊𝒏𝒄 Example: 𝟏𝟎𝟎. 𝟎𝟎𝟎 ∙ 𝟓 − 𝟑. 𝟓 ∙ 𝟏% ∙ 𝟒𝟎𝟎€ = 600.000 € Propensity-to-buy monetization example Sample of the whole client base baseline
- 37. 37 ~400 000 clients Case Study: Predicting consumer loans Propensity score of the client base Histogram X – Predicted probability (next 12 months) Y – Number of clients Logistic transformation - SCORE Histogram X – SCORE Y – Number of clients
- 38. 38 TARGET Loans taken by the client base Performance: Overall PREDICTION ACCURACY AUC = 0,869 Gini = 0,738 K-S = 0,586
- 39. 39 Performance: Quantiles 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Relative cumulative frequency (whole base) Loan acceptance period (months) Clients with the highest SCORE (top 20%) took out more loans in 60 days than the rest of the clients combined within 1 year.
- 40. 40 Clients’ SCORE vs. banking segmentation by activity (CLN_SEG) Client segments TARGET rate CLN_SEG ACT SAC NEW NAC SAV CLOSED TOTAL PROP_12M_20TILE 1 0,20% 0,10% 0,00% 0,10% 0,00% 0,10% 0,10% 2 0,40% 0,40% 0,00% 0,30% 0,10% 0,10% 0,30% 3 0,60% 0,30% 0,00% 0,50% 0,40% 0,20% 0,50% 4 0,70% 0,40% 0,80% 0,60% 1,50% 0,30% 0,60% 5 1,10% 1,10% 1,80% 0,80% 0,50% 0,50% 0,90% 6 1,40% 1,10% 0,50% 0,80% 2,50% 0,30% 1,10% 7 1,90% 1,40% 1,00% 1,50% 0,00% 0,30% 1,60% 8 2,10% 2,00% 2,40% 1,40% 0,00% 0,50% 1,80% 9 2,70% 3,30% 1,60% 1,70% 2,10% 0,70% 2,30% 10 3,60% 2,80% 3,00% 1,80% 0,00% 0,80% 2,90% 11 4,30% 3,50% 4,50% 2,20% 0,00% 1,10% 3,50% 12 5,80% 5,40% 4,40% 2,40% 5,00% 1,60% 4,30% 13 6,70% 6,10% 5,60% 3,80% 7,30% 2,30% 5,20% 14 8,00% 9,50% 6,90% 6,70% 4,30% 2,20% 7,10% 15 10,00% 11,60% 9,80% 9,10% 4,30% 2,90% 9,40% 16 14,40% 16,70% 14,10% 12,00% 11,10% 3,80% 13,10% 17 19,50% 21,50% 19,50% 15,30% 0,00% 3,80% 16,90% 18 27,80% 28,30% 25,10% 20,40% 11,10% 4,70% 25,20% 19 38,50% 35,20% 35,50% 26,10% 0,00% 9,20% 35,90% 20 57,60% 48,00% 48,80% 32,70% 0,00% 12,00% 54,60% TOTAL 13,70% 11,40% 11,20% 4,70% 0,40% 1,50% 9,40% TOP 10% SCORE
- 41. 41 Case Study - SUMMARY TARGET rate CNT_CL_EVER 0 1 [2; 3] [4; 7] [8; 15] >= 16 CLN_SEG ACT 4,8% 24,8% 39,2% 54,6% 73,9% 92,7% SAC 2,6% 18,9% 31,0% 44,9% 63,3% NEW 5,5% 27,8% 57,9% NAC 1,2% 9,1% 17,7% 27,4% 41,1% SAV 0,3% 3,2% 9,8% CLOSED 0,8% 2,7% 4,7% 8,6% % CLN Target rate AUC 31% 1,2% 0,75 1% 70,9% 0,59 68% 12,4% 0,84 Inactive & Savers AUC .75 Frequent borrowers AUC .59 Active & Newcomers AUC .84 31 % of clients 68 % of clients 1 % of clients
- 42. 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 Find out more at bigdataforbanking.com Thank you for joining! contact directly: Lukas Dvorak Business Development Manager lukas.dvorak@profinit.eu We need your help to improve! Since you are here, please help us improve our events and webinars by completing our short survey. We appreciate your interest in helping us grow.