Case-based Recommender Systems for Personalized Finance Advisory

Cataldo Musto, Giovanni Semeraro
Case-based Recommender Systems
for Personalized Finance Advisory
Graz (Austria) - 16.04.2015

one minute
on the Web
C.Musto, G.Semeraro - Case-based Recommender Systems for Personalized Finance Advisory
FinRec 2015 - 1st International Workshop on Personalization and Recommender Systems in Financial Services - Graz (Austria) - 16.04.2015

we can handle 126 bits of information
we deal with 393 bits of information
ratio: more than 3x(Source: Adrian C.Ott,The 24-hour customer)

(from Matrix)
decision-making
is actually challenging

paradox of choice
(Barry Schwartz,TED talk “Why more is less”)

(ﬁnancial) overload

solution: personalization

to adapt asset
portfolios
on the ground of personal
user proﬁle and needs
Insight:

Solution
Recommender Systems

Recommender Systems
Relevant items (movies, news, books, etc.) are suggested to
the user according to her preferences.

deﬁnition
Recommender Systems have the goal of guiding the
users in a personalized way to interesting
or useful objects in a large space of possible
options.
Burke, 2002 (*)
(*) Robin D. Burke: Hybrid Recommender
Systems: Survey and Experiments. UMUAI,
volume 12, issue 4, 331-370 (2002)

does it ﬁt our scenario?
“we are leaving the age of information, we are entering the age of recommendation”
(C.Anderson,The LongTail.Wired. October 2004)

Recommender Systems
“[...] The technology is used by shopping websites such as Amazon,
which receives about 35 percent of its revenue via product
recommendations. It is also used by coupon sites like Groupon; by
travel sites to suggest flights, hotels, and rental cars; by social-
networking sites such as LinkedIn; by video sites like Netflix to
recommend movies and TV shows, and by music, news, and food
sites to suggest songs, news stories, and restaurants, respectively.
Even financial-services firms recently began using
recommender systems to provide alerts for investors about
key market events in which they might be interested”
(N.Leavitt,“A technology that comes highly recommended” - http://tinyurl.com/d5y5hyl)

Recommender Systems
ﬁnancial services
http://www.bloomberg.com/news/articles/2015-03-16/smart-beta-etfs-attract-billions-with-critics-blaming-dumb-money

Recommender Systems
success stories
“People who bought…”
on Amazon
“Discover”
on Spotify

Recommender SystemsRecommender Systems
unexpected stories

recommending ﬁnancial products
is a complex task

ﬂocking

ﬂocking
Too many users could be moved
towards the same suggestions

ﬂocking
consequence: price manipulation
(as in trader forums)

poor knowledge

Features describing both assets
classes and private investors are
poorly meaningful
poor knowledge

poor history
A combination of asset classes
is typically kept for a long time

Solution
Case-based Recommender Systems

case-based RSs
• Inspired by case-based reasoning
• Similar problems solved in the past are
used as knowledge base
• Reasoning by analogy
• The recommendation process relies on
the retrieval and the adaptation of the
solutions adopted to solve similar cases

....but
what do we actually mean with ‘case’ ?

case base
• A case is a the formalization of a
previously solved problem
• In our setting
• Description of a user
• Description of a portfolio
• An evaluation of the proposed solution

case-base
example
user solution evaluation

case-base
example
User Features
Risk Proﬁle: Low
Financial Experience: High
Financial Situation:Very High
Investment Goals: Medium
Temporal Goals: Medium

case-base
example
User Features
Risk Proﬁle: Low
Euro Bonds 30%
High-Yield Bonds 10%
Fixed-Rate bonds 22%
Euro Stocks 23%
Emerging Market Stocks 7%
Money Market 8%

case-base
example
User Features
Risk Proﬁle: Low
monthly rate (e.g.)
+0.22%
Euro Bond 30%
High-Yield Bonds 10%
Fixed-Rate bonds 22%
Euro Stocks 23%
Emerging Markets Stocks 7%
Money Market 8%

case-based RSs
solving cycle

case-based reasoning for
personalized wealth management

scenario
“Scrooge McDuck wants to
get richer. He decided to
invest some of his savings
and he asked for help to a
ﬁnancial advisor”

step 1
user modeling

scenario
Which features
may describe
Scrooge McDuck?
step 1

scenario
User Features
Risk Proﬁle: Low
Investment Horizon High
Investment ExperienceVery High
Financial Assets: Medium
step 1

User Features
Risk Proﬁle: Low
scenario
MiFID-based
step 1

scenario
step 1
+
Generic Demographical Features
User Features
Risk Proﬁle: Low

in a classical pipeline, the target user
would have received a “model” portfolio
tailored on her proﬁle

in a pipeline fostered by a recommender system, the ﬁnancial
advisor can analyze the portfolios proposed to similar users
to tailor the proposal

step 2
neighbors identiﬁcation

given a case base, it is necessary to
deﬁne a similarity measure to
compute how similar two cases are

trivial similarity: user match
two cases are similar if they share
exactly the same features

cases are represented
as points in a vector space
geometrical alternative: cosine similarity

geometrical representation

case-based RSs
each case is seen as a vector

case-based RSs
calculation over the n features

case-based RSs
calculation over the n features
= (risk proﬁle, experience, goals, etc.)

case-based RSs
inner product

case-based RSs
it returns the cosine of the angle
between A and B

case-based RSs
case_A
case_B
cosine

scenario
case base
step 2

scenario
step 2
0.3
0.7
0.9
0.1
similarity score

scenario
step 2
0.3
0.7
0.9
0.1
neighborhood
(helpful cases)

step 3
extraction of candidate portfolios

scenario
Euro Bonds 30%
HighYield Bonds 15%
Fixed Rate Bonds 15%
Europe Stocks 20%
Money Market 8%
Euro Bonds 30%
HighYield Bonds 10%
Europe Stocks 23%
Flessibili BassaVolatilità 8%
step 2
solutions proposed to the neighbors
are labeled as candidate solutions

step 4
ranking of candidate portfolios

in real-world scenarios, the case base
contains many helpful cases

in real-world scenarios, the case base
contains many helpful cases
it is necessary to introduce strategies
to ﬁlter and rank the cases

revise
We implemented several ranking strategies
• Temporal ranking
• Clustering
• Diversiﬁcation
• Financial ConﬁdenceValue (FCV)

revise
temporal ranking
solutions are ranked from the newest to the oldest (or viceversa)

Euro Bonds 30%
HighYield Bonds 15%
Europe Stocks 20%
Money Market 8%
Euro Bonds 30%
HighYield Bonds 10%
Europe Stocks 23%
Money Market 8%
Euro Bonds 15%
HighYield Bonds 25%
Europe Stocks 40%
Money Market 8%
Euro Bonds 20%
HighYield Bonds 20%
Europe Stocks 35%
Money Market 8%
revise
temporal ranking

Euro Bonds 30%
HighYield Bonds 15%
Europe Stocks 20%
Money Market 8%
olderolder
Euro Bonds 30%
HighYield Bonds 10%
Europe Stocks 23%
Money Market 8%
Euro Bonds 15%
HighYield Bonds 25%
Europe Stocks 40%
Money Market 8%
Euro Bonds 20%
HighYield Bonds 20%
Europe Stocks 35%
Money Market 8%
revise
temporal ranking

clustering
solutions are clustered and just a small set of centroids is proposed
revise

clustering
revise
Euro Bonds 30%
HighYield Bonds 15%
Europe Stocks 20%
Money Market 8%
Euro Bonds 30%
HighYield Bonds 10%
Europe Stocks 23%
Money Market 8%
Euro Bonds 15%
HighYield Bonds 25%
Europe Stocks 40%
Money Market 8%
Euro Bonds 20%
HighYield Bonds 20%
Europe Stocks 35%
Money Market 8%

clusteringcluster 1
revise
Euro Bonds 30%
HighYield Bonds 15%
Europe Stocks 20%
Money Market 8%
Euro Bonds 30%
HighYield Bonds 10%
Europe Stocks 23%
Money Market 8%
Euro Bonds 15%
HighYield Bonds 25%
Europe Stocks 40%
Money Market 8%
Euro Bonds 20%
HighYield Bonds 20%
Europe Stocks 35%
Money Market 8%

clusteringcluster 1 cluster 2
revise
Euro Bonds 30%
HighYield Bonds 15%
Europe Stocks 20%
Money Market 8%
Euro Bonds 30%
HighYield Bonds 10%
Europe Stocks 23%
Money Market 8%
Euro Bonds 15%
HighYield Bonds 25%
Europe Stocks 40%
Money Market 8%
Euro Bonds 20%
HighYield Bonds 20%
Europe Stocks 35%
Money Market 8%

insight: ﬁltering out too similar solutions
diversiﬁcation algorithm
revise

revise
identiﬁcation of the best subset of similar cases
which maximize the relative diversity

revise

revise
input
similar cases
(candidate solutions)

revise
output
subset of
diversiﬁed cases

revise
algorithm
in each step the
portfolio which
best diversiﬁes
the solutions is
chosen

revise
Solutions with
the highest
quality are
iteratively
chosen

revise
combination
between
similarity and
diversity

it returns portfolios
that are not so
similar to those
previously put in the
result set
revise

revise
Euro Bonds 30%
HighYield Bonds 15%
Europe Stocks 20%
Money Market 8%
Euro Bonds 30%
HighYield Bonds 10%
Europe Stocks 23%
Money Market 8%
Euro Bonds 15%
HighYield Bonds 25%
Europe Stocks 40%
Money Market 8%
Euro Bonds 20%
HighYield Bonds 20%
Europe Stocks 35%
Money Market 8%

revise
Euro Bonds 30%
HighYield Bonds 15%
Europe Stocks 20%
Money Market 8%
Euro Bonds 30%
HighYield Bonds 10%
Europe Stocks 23%
Money Market 8%
Euro Bonds 15%
HighYield Bonds 25%
Europe Stocks 40%
Money Market 8%
Euro Bonds 20%
HighYield Bonds 20%
Europe Stocks 35%
Money Market 8%
X
X

revise
Financial ConﬁdenceValue (FCV)
• Simple insight
• We know the historical yield for each of
the assets class in the portfolio
• FCV ranks ﬁrst the solutions composed
by a combination of asset classes close
to the optimal one (according to
previous yield)

revise
(Generated yield) (Drift Factor)Total yield is the
product of the
yield generated
by each asset
class with the its
percentage in the
portfolio
Ratio between
the yield
generated by the
asset classes in
the portfolio and
its complement

revise
Euro Bonds - - - 30%
HighYield Bonds 15%
Europe Stocks +++ 20%
Money Market 8%
HighYield Bonds 10%
Money Market 8%
HighYield Bonds 25%
Money Market 8%
HighYield Bonds 20%
Money Market 8%

step 5
discussion of the solution

ﬁnancial advisor and private investor
can further discuss the portfolio

review
Original Discussed Gap
Euro Bonds 30% 30%
HighYield Bonds 12.5% 10% -2.5%
Fixed Rate Bonds 18.5% 20% +1.5%
Europe Stocks 21.5% 24% +2.5%
Emerging Markets
Stocks 9.5% 8% -1.5%
Money Market 8% 8%
interactive personalization

step 6
case base update

an evaluation score is ﬁnally assigned to the proposed
solution
yield, e.g.
retain
good solutions are stored in the case base and exploited
for future recommendations

case base

(new) case base

our implementation

our implementation
http://193.204.187.192:8080/OBWFinance
demo available

OBWFinance
login screen
advisor-oriented tool

OBWFinance
client selection

OBWFinance
recommendation parameters

OBWFinance
only admins can change the parameters

OBWFinance
one click to generate recommendations

OBWFinance
drop-down menu for selecting the best solution

OBWFinance
assets class

OBWFinance
yield of the solution

OBWFinance
chosen portfolio can be further discussed

evaluation

evaluation
what is the average yield of
recommended portfolios?

evaluation
what is the average yield of
recommended portfolios?
can recommender systems suggest
better investment portfolios than
human advisors?

design of the experiment
• 1172 users
• 19 assets classes
• Different neighborhood sizes
• Different features describing the users
• Risk Proﬁle, Investment Goals, Investment Horizon,
Investment Experience, Financial Assets, Advice Type, Sex,Age
• Different similarity measures (Cosine vs. UserMatch)
• Leave-one-out experimental design
evaluation

experiment 1
user match vs. cosine similarity
Yield
0
0,04
0,08
0,12
0,16
0,2
neighbors
1 5 10
0,2
0,19
0,18
0,1
0,11
0,09
User Match Cosine Sim
cosine similarity overcomes user match

experiment 2
how many features?
Yield
0
0,042
0,084
0,126
0,168
0,21
neighbors
1 5 10
0,2
0,21
0,2 0,2
0,19
0,18
Financial Features Financial + Demographical Features
cosine similarity overcomes user match

experiment 3
revise strategies (yield)
best performing conﬁguration provides 0,28% monthly yield
Yield
0
0,056
0,112
0,168
0,224
0,28
neighbors
1 5 10
0,250,24
0,22
0,270,28
0,22
0,2
0,15
0,13 0,14
0,12
0,09
0,20,210,2
Basic Clustering Diversiﬁcation FCV FCV + Div

experiment 3
revise strategies (diversity of the solutions)
ILD=1-average similarity between portfolios
Intra-ListDiversity(ILD)
0
0,14
0,28
0,42
0,56
0,7
neighbors
0,58
0,35
0,7
0,46
0,41
Basic Clustering Diversiﬁcation FCV FCV + Div

experiment 4
comparison to baselines (leave-one-out evaluation)
recsys better than humans!
Yield
0
0,056
0,112
0,168
0,224
0,28
neighbors
1 5 10
0,270,28
0,22
0,20,20,2
0,170,170,17
Human Collaborative FCV

• FCV calculated on January, 2014
• Recommendations generated on January, 2014
• Evaluation of the yield generated from
February 2014 to July 2014
experiment 5
ex-post evaluation (6 months, with real data)

experiment 5
ex-post evaluation (6 months, with real data)
FCV and Diversiﬁcation is the best one
Yield
0
0,032
0,064
0,096
0,128
0,16
neighbors
1 5 10
0,060,060,06
0,040,04
0,05
0,11
0,12
0,16
0,09
0,1
0,16
0,06
0,08
0,15
Basic FCV FCV + Div Collaborative Human

•Personalized Wealth Management
• Application of case-based reasoning
• Geometrical similarity measure to identify the most
similar previously solved cases
• Introduction of diversiﬁcation and re-ranking
techniques
• More than 3% yield for year
• Experiments shows that recommended portfolios
overcome the real ones for almost all the users
• Working Demo!
recap

questions?
Giovanni Semeraro
giovanni.semeraro@uniba.it
Cataldo Musto
cataldo.musto@uniba.it
in memoriam
Aaron Swartz

Case-based Recommender Systems for Personalized Finance Advisory

Recommended

Recommended

More Related Content

Viewers also liked

Viewers also liked (20)

Similar to Case-based Recommender Systems for Personalized Finance Advisory

Similar to Case-based Recommender Systems for Personalized Finance Advisory (20)

More from Cataldo Musto

More from Cataldo Musto (20)

Recently uploaded

Recently uploaded (20)

Case-based Recommender Systems for Personalized Finance Advisory