recommender systems for mass customization of financial advice · recommender systems for mass...
TRANSCRIPT
BRANDSCHENKESTRASSE 41
CH-8002 ZURICH
SWITZERLAND
INCUBEGROUP.COM
Recommender Systemsfor Mass Customizationof Financial Advice
SDS|2018, Bern, 07.06.2018
Anna Nowakowska
Head of Data Analytics
2
1
2
3
Data Analytics Team at InCube
Recommender Systems for Financial Advice
Private Banking Use Case
4 Retail Banking Use Case
5 Outlook for the Future and Summary
AgendaTalk Outline
3
About InCubeOur Focus
We develop fintech products. We provide consulting services.
Full stack product development and expertise-driven consulting
Products and Services for Financial Institutions
Wealth
Tech
Full stack solutions for digital
wealth management driven by
algorithms, artificial intelligence
and user experience
Data
Science
Artificial intelligence for real-word
use cases at financial institutions –
from rapid prototyping to scalable
production systems
Quantitative
Finance
Robotic process automation and
digitization require scalable and
robust quant solutions for
investment, risk and compliance
management
Business
Consulting
Consulting services from project
management, business analysis
and transformation, subject matter
expertise to implementation
4
Data
analytics
Software
Engineering
Business
Consulting
Quantitative
Finance
Backend
Engineering
Frontend
Web Apps
Regulatory
Compliance
Business
Analysis
& PM
Machine Learning
Reporting &
Visualization
Predictive
Analytics
Architecture
& Design
UX & UI
Design
InCube Services Data Analytics Team
Data Analytics Team at InCubeIntroduction
Machine
Learning
Anna NowakowskaMEng Electronics & Electrical Engineering, CFA
Head of Data Analytics
Coordinating the recommender system projects
Milica PetrovićMSc ETH Statistics
Data Scientist
Feature-based recommender for retail banking products
Aleksandra ChirkinaMSc ETH Statistics
Data Scientist
Feature-based recommender for retail banking products
Jeremy SeowETH Statistics Master’s Student
Data Science Intern
Advanced collaborative filtering recommender for private
banking products
5
Recommender Systems at InCubeTrack Record
Recommender Systems
2016-2017
Stock Recommendation Application
▪ Stock recommendations based on factor selection and weighting
▪ Factor selections can be self-configured or learned by data
H1 2017
Case-Based Reasoning Techniques
H2 2017
Collaborative Filtering Techniques
Q1 2018
Model Fitting to Data of Pilot Bank
Q2 2018 -
Model Extensions with Our Clients
▪ Portfolio proposals based on case base library
▪ Client similarity measures, portfolio retrieval and ranking algorithms tested
▪ Started to cooperate with ETH Zurich on recommender systems
▪ Fitted and tested various collaborative filtering (CF) and boosting techniques
▪ Model- and neighbourhood-based CF models with Nidwaldner Kantonalbank
▪ Working on use cases for retail and private banking
▪ Started projects with two large banks in CH for an advisory recommender
▪ Combine (deep) autoencoders with CF to learn feature representations for
latent factor models in CF
6
1
2
3
Data Analytics Team at InCube
Recommender Systems for Financial Advice
Private Banking Use Case
4 Retail Banking Use Case
5 Outlook for the Future and Summary
7
Optimization engine Optimal portfolio
OutputsOptimizationInputs
Assets
Risk profile
Constraints
Objectives
▪ Many portfolio management processes incl. robo-advisors are based on Modern Portfolio Theory (MPT)
▪ These tools maximize the client’s utility function based on a client risk profile and an applicable offering
▪ However, client’s affinity to the proposed portfolio and products is usually not considered
Recommender Systems for Financial AdviceMaximizing the Client’s Utility Function
Cumulated returns of different securities Efficient FrontierCumulated returns of different securities Weights
Target Risk[Cov]
Valu
e
8
Business Value
▪ Maximize client’s utility function and at the same time
the acceptance rate of proposals
▪ New opportunities and reduced manual work for the
wealth managers for HNWI and UHNWI client
segments
▪ Automated but customized advice for retail client
segments
Recommender Systems for Financial AdviceMaximizing the Acceptance Rate of Proposals
Problems
▪ Huge amount of data
▪ 50 to 200 clients per relationship manager
▪ Hard to manage manually
Solution
▪ Similar clients – similar financial needs
▪ Many financial products share similar characteristics
▪ Recommender systems:
➔ propose products used by similar clients
➔ propose similar but unseen products to clients
➔ improve and scale-out customized financial advice
"People who bought this also liked…"
"If you bought this, you might also like…"
9
1
2
3
Data Analytics Team at InCube
Recommender Systems for Financial Advice
Private Banking Use Case
4 Retail Banking Use Case
5 Outlook for the Future and Summary
10
Private Banking Use CaseNature of the Use Case and Data Description
Subset of data from Nidwaldner Kantonalbank
✓ provided by the Nidwaldner Kantonalbank (NKB) and
anonymised for the purpose of this presentation
✓ 1117 users, 1788 items
✓ ~18 products per person (average)
Many offered products (thousands; financial instruments such as stocks, bonds, derivative instruments, etc)
Clients typically own many products
Clients buy and sell investments - substantial history of user-product “ratings”
Goal: personalized recommendations of financial instruments, such as stocks and bonds
11
Private Banking Use CaseNKB Data Overview: Ratings Matrix
• Two groups of clients visible
• 1st group: simple behavior
• 2nd group: complex behavior
• |2nd group| > |1st group|
• Complex data structure
→ we expect a popular model to not
be good enough here
12
Private Banking Use CaseNKB Data Overview: How Many Products per Client?
Number of products held by client
% o
f cl
ien
ts
Clients hold more than a few products
Average number of 18 products per client, which is a good
basis for using collaborative filtering models
18
13
Private Banking Use CaseCollaborative Filtering Approach
Model-based collaborative filtering (CF)
+ no user or item features required
+ typically more accurate than other models
- cold start
- difficult to interpret
→useful for private
banking use case
(abundant history)
Challenges
missing data points
→ client didn’t want the product? OR doesn’t know about it?
implicit ratings
→ how to determine if the clients liked the products they bought?
Modeling approach
1. Ratings matrix factorization to discover latent features
2. Confidence weights – fix confidence weights in one model
3. Boosting – confidence weights estimated from the ensemble model
14
Private Banking Use CaseResults
RESULTS
Model-based CF Popular Model
AUC 0.9077 ± 0.0003 0.8728
nDCG 0.3586 ± 0.0004 0.2656
Model based CF wins against the popular model
15
1
2
3
Data Analytics Team at InCube
Recommender Systems for Financial Advice
Private Banking Use Case
4 Retail Banking Use Case
5 Outlook for the Future and Summary
16
Retail Banking Recommender DemoQlik Sense App: Top 5 Account Recommendations per Client
17
Retail Banking Recommender DemoQlik Sense App: Account Recommendations Across Client Segments
18
Retail Banking Use CaseNature of the Use Case and Data Description
Data from Nidwaldner Kantonalbank
✓ provided by the Nidwaldner Kantonalbank (NKB) and
anonymised for the purpose of this presentation
✓ ~100 available products
✓ Product categories:
• Current accounts
• Savings accounts
• Credit card accounts
• Investment accounts
• Mortgages and loans
• etc.
Relatively few offered products (tens to hundreds; current or savings accounts, credit cards, etc.)
Clients typically own few products
Clients rarely change products
Goal: personalized recommendations of retail banking products, such as accounts and credit cards
19
Retail Banking Use CaseData Overview: Top 20 Most Popular Products and % of Clients Holding Them
pro
du
ct
% of clients holding the product
A few very popular products
A few products popular across the client base,
the rest only held by smaller groups of clients
20
Retail Banking Use CaseData Overview: How Many Products per Client?
Most clients only hold a few accounts
Average number of 2.9 products per client, over the entire
history of the client being with the bank
% o
f cl
ien
ts
Number of products held by client2.9
21
Retail Banking Use CaseWhy Model Based Collaborative Filtering Fails & Our Model Choice
RESULTS
Model-based CF Popular Model
Accuracy 8.96 ± 0.18% 20.43 ± 0.23%
Mean Reciprocal Rank 17.72 ± 0.18% 31.53 ± 0.22%
“Popular” (non-personalised) model performs better than model-based CF
• clients consume too few products
• low variety of the most popular products
Memory-based demographic collaborative filtering
+ cold-start solved through user features
+ easily interpretable: explanations for recommendations
- requires collection of features
- need to store full matrix
→ useful for retail banking use case (limited history)
22
Retail Banking Use CaseMemory Based Demographic Collaborative Filtering: Approach
Step 1: Demographic segmentation
For each client, find a neighbourhood of k similar clients (k-NN)
based on the Gower distance and features:
• gender
• age group
• wealth group
• e-banking usage
• 3rd pillar payments
• ...
Step 2: Product popularity
Within each client’s neighbourhood:
• determine the popularity of each product - how many clients in the
neighbourhood consumed it
Step 3: Personalized recommendation
For each client:
• sort the products the client has not yet consumed by popularity in
his / her neighborhood
• recommend the top 5 products
23
Retail Banking Use CaseResults
RESULTS
Model-Based CF Popular Model Memory-Based Demographic CF
Accuracy 8.96 ± 0.18% 20.43 ± 0.23% 45.11 ± 1.27%
Mean Reciprocal Rank 17.72 ± 0.18% 31.53 ± 0.22% 58.44 ± 1.01%
Memory based demographic CF wins against popular model and model-based CF
24
1
2
3
Data Analytics Team at InCube
Recommender Systems for Financial Advice
Private Banking Use Case
4 Retail Banking Use Case
5 Outlook for the Future and Summary
25
Outlook for the futureRecommender Systems for Mass Customization of Financial Advice
Ongoing projects
✓ Working with two large Swiss banks on an advisory
recommender system in private banking
Modeling and evaluation approach
✓ Online evaluation → explicit feedback
✓ A/B testing
✓ Features changing in time
✓ Hybrid models
✓ Incorporating both item and client features
26
SummaryRecommender Systems for Mass Customization of Financial Advice
Private Banking: personalized recommendations of financial instruments
• Many products, diversity of investing behaviour across the client base, abundant history of ratings per user
• Model-based Collaborative Filtering
• higher confidence weights for non-missing data points
• start with a simple matrix decomposition model with fixed confidence weights
• apply boosting with a number of component models, between which confidence levels are adjusted
Retail Banking: personalized recommendations of accounts, credit cards, mortgages, etc.
• Few products per user, limited history of ratings
• similarities between users calculated using the Gower distance
• for each user, his “neighbourhood” of 150 most similar users is determined
• the most popular products in this neighbourhood are recommended to this user
• Not suitable for ratings matrix decomposition
• Memory-based Demographic Collaborative Filtering
BRANDSCHENKESTRASSE 41
CH-8002 ZURICH
SWITZERLAND
INCUBEGROUP.COM
Thank you!
Anna Nowakowska