Using new advances in machine learning, we tackled 4 pain points of the insurance industry: Accelerated Underwriting process, Personalized Pricing, Fraud Detection & Improved Customer Retention and Experience
1. KOI OS
P R E S E N T A T I O N
Adding Value to the Insurance Industry
Intelligence
2. 1 NEW DEVELOPMENTS IN THE INSURANCE INDUSTRY
2 PROCESS AUTOMATION
3 PERSONALIZED PRICING
4 RISK MANAGEMENT
5 TEAM
TABLE OF CONTENTS
3. 1PART
NEW DEVELOPMENTS IN THE INSURANCE INDUSTRY
PROCESS AUTOMATION
PERSONALIZED PRICING
RISK MANAGEMENT
CUSTOMER EXPERIENCE
OUR TEAM
4. Through the unlocking of potential in these 4 areas!
PROCESS
AUTOMATION
PERSONALIZED
PRICING
ADAPTED RISK
MANAGEMENT
CUSTOMER
EXPERIENCE
01!Koïos Intelligence Inc.!
NEW ADVANCES IN BIG DATA AND MACHINE LEARNING
ARE DISRUPTING THE INSURANCE INDUSTRY
5. 02!Koïos Intelligence Inc.!
UNLOCKING VALUE THROUGH PROCESS AUTOMATION !
OUTDATED MODEL NEW DEVELOPMENTS EVOLVED MODEL
02 !01!
• Insurance claims and
property valuations taking a
long time to be processed
and involving too many
different staff members,
resulting in high labour costs
and poor customer
satisfaction!
• Optical character recognition
technology to allow fast
transformation of paper
documents into computer
readable format!
• Improved data analysis
using new machine learning
models!
!
!
• Improved productivity and
reduced costs for complex
processes such as property
valuations
• Claims processed faster
through automated image
classification
6. 03!Koïos Intelligence Inc.!
UNLOCKING VALUE THROUGH PERSONALIZED PRICING !
OUTDATED MODEL NEW DEVELOPMENTS EVOLVED MODEL
02!01!
• General price segmentation
resulting in undeserved
customers who do not fit in
general categories
• Product recommendations
based on advisor’s
experience
• Big amounts of personal data
collected though
smartphones and other
devices (e.g. health, lifestyle,
driving behaviour, etc.)
• New predictive analysis tools
use new variables (e.g.
security, traveling, multiple
coverage, etc.) allowing more
accurate pricing than before
• Recommendation of more
personalized products
resulting from in-depth
analysis of customer data
• Novel personalized pricing
using individual risk profile
7. 04!Koïos Intelligence Inc.!
UNLOCKING VALUE THROUGH RISK MANAGEMENT !
OUTDATED MODEL NEW DEVELOPMENTS EVOLVED MODEL
02!01!
• Risk groups classified using
static data from very few
metrics(e.g. age, sex, and
medical history)
• Many flaws and false alarms
in insurance fraud detection
• The advent of artificial
intelligence is enabling the
development of extremely high-
capacity models that can
analyze previously unimaginable
amounts of data, that are
extremely powerful in a fraud
detection context
• Insurance fraud detected more
accurately using holistic
customer analysis
• Significant reduction of false
positives
• Risk group classified using a
wide spectrum of data such as
social networks, clickstreams
and web analytics
8. 05!Koïos Intelligence Inc.!
UNLOCKING VALUE THROUGH CUSTOMER EXPERIENCE !
OUTDATED MODEL NEW DEVELOPMENTS EVOLVED MODEL
02!01!
• Inconsistent customer service
quality, dependant on agent
experience and skill
• Customer experience analysis
after losing the client
• Churns (decommissioning rate)
customer predicted and
intercepted in real time
• Proactive customer service by
identifying new needs during
important life events of
customers. Following up on
quotes and contract
terminations
• Robust automated learning
models that ensure constant
reliability in spite of
modifications in the data
• These models allow the precise
interpretation of client
responses in real time
9. 2PART
NEW DEVELOPMENTS IN THE INSURANCE INDUSTRY
PROCESS AUTOMATION
PERSONALIZED PRICING
RISK MANAGEMENT
CUSTOMER EXPERIENCE
OUR TEAM
10. Analyze customer’s profile and assess customer’s need!
CHATBOT!
Make a Proposal (offer/quote)!
If quote is
approved!
Resiliate the contract!
Retract (cancel)!
Modify the personal
situation!
Request documents from
client!
Before Sale! Sale! After Sale!
Proposal! Smart contract!
Adjust quote!
Modifications / Endorsements!
PERSONALIZED
PRICING
CLIENT
EXPERIENCE &
COMPENSATION
RISKMANAGEMENT
PROCESS AUTOMATION
AUTOMATION OF THE SUBSCRIPTION PROCESS
06!Koïos Intelligence Inc.!
Subscribe a
contract!
Modify the
contract!
11. 1!
Digitization is the process of transforming forms and handwritten documents into a computer-readable format using machine learning
algorithms!
Digitization !
Fraud
Detection!
Manual
Work!
Online
Marketing! Call
Centres!
Automatization of Manual Work!
Reduce the number of errors made by
manual labor!
Claims Processing!
Process claims and automate changes
and simple queries (virtual agents)!
Fraud detection!
Digitization is the first step in collecting
textual data that can improve fraud
detection models!
Acquiring customers online!
Digitization enables an end-to-end online
distribution channel through targeted
generation of leads on social media!
07!Koïos Intelligence Inc.!
PROCESS AUTOMATION
DIGITAZION APPLICATIONS
12. 08!Koïos Intelligence Inc.!
Classification! FAQ!
1
(Future work)
Conversion of speech
to text
2
Classification of the user’s
message in predetermined
categories (e.g. date, amount
$, etc.)
3
Searching the DB allows to
determine if the category found
is adequate knowing the
previous question
4
If needed, the answer is
stocked in a DB
5
If the answer is adequate then
the chatbot can send the next
question
6
Otherwise the message is sent to
a second algorithm which will
send an answer based on
predefined questions
PROCESS AUTOMATION
CHATBOT
13. From the profile of a client and his claim history, the pricing process begins with a model that identifies factors with a high predictive
potential to then build a description of severity as a function of the characteristics of the clients!
09!Koïos Intelligence Inc.!
Prediction!
Coverage!
Vehicle!
Profile!
!Vehicle!
• Brand of the car!
• Weight!
• Type of fuel!
!Profile!
• Age!
• Civil status!
• Number of drivers!
!Coverage!
• Frequency of payments!
• Term!
PROCESS AUTOMATION
AUTOMATED CLAIM SEVERITY ASSESSMENT
14. Using the data of 180,000 claims, we designed and tested a model to predict the claim severity from 116 categorical factors (e.g. car brand)
and 14 continuous factors (e.g. volume of the car’s engine)!
Optimal predictions are reached by combining the
performance of all models. Value of claims predicted
with a mean absolute error of 1110.01362!
Fig. 3 Architecture of a model of claim
severity prediction!
The mean absolute error is a simple and commonly used measure that indicates the spread between predicted and observed values. It is the
arithmetic average of the absolute value of the spreads!
10!Koïos Intelligence Inc.!
PROCESS AUTOMATION
AUTOMATED CLAIM SEVERITY ASSESSMENT
15. 3PART
NEW DEVELOPMENTS IN THE INSURANCE INDUSTRY
PROCESS AUTOMATION
PERSONALIZED PRICING
RISK MANAGEMENT
CUSTOMER EXPERIENCE
OUR TEAM
16. Customized services are becoming increasingly popular and have the potential to create huge, high-quality data streams, increase the
frequency of client-insurer interactions, and significantly reduce the number of claims by improving health, safety, and the security of the
insured!
INSURANCE
INDUSTRY!
Car! House!
Health! Life!
Safe Driving & Support!
• Alerts and rewards for safe driving!
• Alerts for theft or damage!
• Troubleshooting service!
• Assistance to find a car!
Healthy Lifestyle!
• Rewards for a healthy lifestyle!
• Health checks and access to
behavioral data!
• Schedule a meeting with a doctor!
• Diagnosis and remote consultation!
Home security!
• Remote monitoring and alerts!
• Automatic shutdowns during water
leakage & faster fire extinguisher settings !
Energy!
• Aiming towards a smart home!
• Tips for saving energy!
Financial Planning!
• Pension and real estate planning!
• Tips on personal finances!
11!Koïos Intelligence Inc.!
PERSONALIZED PRICING
INSURANCE PERSONALIZATION
17. The Internet offers insurers a holistic view of their policyholders, improving the quality of their models and enabling them to offer
personalized products!
Car!
House!
Health!
Life!
Constant flow of
consumer data!
Personalized
insurance
products!
Score
Adjustment!
✓ New markets !
✓ Improved models !
✓ Better client
retention
!
Client Score!
Sharing data with
the ecosystem!
12!Koïos Intelligence Inc.!
PERSONALIZED PRICING
INSURANCE PERSONALIZATION
18. SLEEP
DIET
PHYSICAL CONDITION
ROUTINE
46%
23%
59%
38%
Don’t leave any data on the table
Smartphones can provide information allowing us to better price a life insurance
premium while improving the customer experience, and encouraging customers to
have a better lifestyle.
13!Koïos Intelligence Inc.!
LIFE INSURANCE
19. 2!
Our Research & Development team designed and tested a model on data containing the classification of more than 80,000 insured persons to predict their risk level, from
130 categorical factors (e.g. medical conditions) or continuous (e.g. BMI)
When the three models are used altogether, we
obtain an accuracy value of 0.67921, measured with
the quadratic weighted Kappa!
Fig. 2 Architecture of the model of risk classification
that an insured person represents!
The quadratic weighted Kappa is a measure of agreement between two assessors, adjusted to take into account the probability to predict correctly, but randomly i.e. without
any good reason to predict correctly; the potential values go from -1 (total disagreement) to 1 (total agreement). A value of zero is expected if all the agreements are due to
randomness
14!Koïos Intelligence Inc.!
PERSONALIZED PRICING
INSURANCE PERSONALIZATION
20. 4PART
NEW DEVELOPMENTS IN THE INSURANCE INDUSTRY
PROCESS AUTOMATION
PERSONALIZED PRICING
RISK MANAGEMENT
CUSTOMER EXPERIENCE
OUR TEAM
21. 15!Koïos Intelligence Inc.!
LEGAL COMPLIANCE
All legislative needs (laws and
regulations): Personal Data
Protection and Electronic Data
Protection Act (PIPEDA) Customer
data management according to the
type of information (e.g. public,
confidential, etc.)
ECONOMIC COMPLIANCE
Remote payment and following
compliance for secure payment
(phone, Internet) SEPA rules
TECHNOLOGICAL COMPLIANCE
The data (servers) must remain in
Canada, since the networks of the
Internet providers are so
interconnected between them, it
would be impossible to prove that
the data are not found in American
territory (so that the American
government can have access
according to the Patriot Act) Koïos
Intelligence chose Amazon.
SOCIAL COMPLIANCE
Anti-Spam Law (abbreviated C-28)
Get the consent of customers
before contacting them, "almonds"
are very very high and especially
cumulative for this law
RISK MANAGEMENT
FACILITATING CONFORMITY AND SECURITY
22. PUBLIC AND
PRIVATE DATA
• Exponential increase in the volume
of reported data
• Increased frequency of quarterly and
annual reports
• Quantitative and Qualitative
disclosures Required
NEW ECOSYSTEM
• Complexity of the group
• New rules and regulations to
comply
• Reconciliations required
• Reliable and accurate data
ACCELERATION
• Limited resources and tight
schedule
• The Pillar III Annual Return will
change from 20 weeks to 14 weeks
• Requires internal planning to deal
with workload fluctuations
INTERSECTION
• Solvency II and IFRS 17 focus on
assessing and managing the risks
facing the insurer and quantifying
future risks
• Assets and liabilities are likely to
use a current valuation method,
which is expected to increase the
volatility of financial statements
relative to current standards
• A best estimate basis is used in
expected future cash flows
• The discount rate used is the sum
of the risk-free rate and the liquidity
premium
Technology!
Technical
Provisions!
SCR / MCR!
Financial steps !
Investment!
Reporting
Xbrl !
0%!
100%!
Large
workload!
t!
Data required
to evaluate:!
MCR, SCR,
ORSA, QRT!
!
Insurance companies are now required to provide more and better information in a shorter period of time on a quarterly and yearly basis. The
Solvency II standard has implemented prudent supervision of both own funds and technical provisions in order to limit the probability of
insolvency of an insurer. IFRS 17 includes the request for financial information relating to an insurer's own funds and technical provisions in
order to assess an overall economic value!
SOLVENCY II / TSAV AND IFRS 17
16!Koïos Intelligence Inc.!
23. ONLINE SERVICES
AUTOMATION
BIG DATA
WHY IT IS TIME TO ACT?
Democratization of internet and overwhelming simplicity to obtain false
identity online triggers an increase in organized frauds
Several repetitive processes such as reading the claim of a client will
largely be simplified and even completely automated, in some cases
We are currently living in the midst of a technological revolution where it is
now possible to analyse any kind of data which were once left to an
expert’s judgement, for instance the images of a car crash
RISK MANAGEMENT
FRAUD DETECTION IN INSURANCE
17!Koïos Intelligence Inc.!
24. CLAIMANT DATA
COVERAGE DATA
DAMAGE DATA
Modelling of variables
First step is to structure the data of the claimant, coverage and damage using different
machine learning algorithms (see example of structured variables below)
Resource requirements
Estimation of resources needed to
launch an investigation for each of the
fraudulent claims
Gain expectation
Estimation of
resources and
probability of fraud
allow us to calculate
the potential gain to
launch an
investigation for a
given claim
Ranking
We can then rank
the claims by order
of priority to
maximize the work
of the investigation
team
Variables
Analysis & processing of data
Coverage
- New coverage
- Profile of the device used
to buy the coverage
- Value of claim
Damage
- Geolocation of the
damage
- Type of damage
- Missing police report
Main Model
Once the data is in a computationally processable form, it can be introduced in the main
model (the main models are presented in the next page. Not that it is possible, even
recommended, to use a set of models)
Claimant
- Important claim history
- Declared revenues does
not correspond objects’
value
Fraud probability
The main model calculates the
probability that the claim is fraudulent
by taking into account all introduced
data
RESULTS
18!Koïos Intelligence Inc.!
RISK MANAGEMENT
FRAUD DETECTION IN INSURANCE CLAIMS
25. ✓ Create an end-to-end system capable of calculating the dollar amount
of the claim using images of a car crash and a minimum of the
information about the car involved in the accident
OBJECTIVES OF THE PROJECT!
• When a claim is received, the images of the accident and the
car data are automatically processed by a deep learning
model that estimates the amount of the claim
• Another model then predicts the probability that this claim is
fraudulent by analyzing the spread between the amount
estimated by the model and the required amount entered by
the client (in practice, it is also useful to transform that
probability in a score)
• It is then possible to incorporate that probability/score in the
main model which will decide, given all the other information
(claimant, coverage), if it is necessary to pursue the
investigation
OURAPPROACH
ILLUSTRATION!
CLIENT BENEFITS!
✓ Allow to incorporate data, which up to now could only be analyzed at the beginning of
an investigation
✓ Enriches the client’s existing models by using another explanatory variable
IMAGE RECOGNITION
19!Koïos Intelligence Inc.!
RISK MANAGEMENT
FRAUD DETECTION IN INSURANCE CLAIMS
Information
processing
Main model to
predict cost of
damages
Comparison of predicted
vs observed values to test
model accuracy
26. Our Research & Development team analyzed the data of 300,000 credit card transactions, 500 of which were fraudulent
SVM Neural Network Auto
Encoder
True positive 68 84 73
False positive 35 2501 953
True negative 49965 47499 49047
False negative 25 9 20
Most of the models besides the SVM resulted largely
in false alarms (false positives)!
Fig. 1 Results on the testing set - 50,093
transactions/103 frauds!
➢ The optimal precision is reached by using an ensemble composed of the models above, but assigning a greater weight to the SVM
to account for the fact that it produces less false alarms than other models
20!Koïos Intelligence Inc.!
RISK MANAGEMENT
FRAUD DETECTION IN CREDIT CARDS
27. 5PART
NOUVEAUTÉS DANS L’UNIVERS DE L’ASSURANCE
ROBOTISATION DES PROCESSUS
TARIFICATION PERSONNALISÉE
GESTION DES RISQUES
EXPÉRIENCE CLIENT
NOTRE ÉQUIPE
NEW DEVELOPMENTS IN THE INSURANCE INDUSTRY
PROCESS AUTOMATION
PERSONALIZED PRICING
RISK MANAGEMENT
CUSTOMER EXPERIENCE
28. CLIENT
EXPERIENCE
1! 2! 3!
4!
5!
6!
7!
1!
2!
3!
4!
5!
6!
7!
Registration in a new!
ecosystem!
Buying a new vehicle – the
offer an insurance product
by the insurer !
The offer of a personalized
life insurance product that
rewards clients with a good
lifestyle !
Receiving an alert during
an abnormal event!
LIFETIME
Alerts sent in case of
damage & vehicle
search by customer
assistance !
Assistance in case of
illness and making
appointments online with a
specialist!
Buying a new house - the
offer of a home insurance
by the insurer!
6!
CUSTOMER EXPERIENCE
BETTER SUPPORT DURING KEY LIFE EVENTS OF CUSTOMERS
21!Koïos Intelligence Inc.!
29. RETENTION
• It is more expensive to acquire a new
client than to retain an existing one
• Clients that renews are the most
profitable on the long run
Key: determine the elasticity of the price in order to
know what increase a client is willing to accept
22!Koïos Intelligence Inc.!
There is no doubt that any change in the pricing has an impact on customer retention. Intuitively, a decrease in pricing is a guarantee of a
high retention rate whereas an important increase will bring more clients to turn to competition.!
CUSTOMER EXPERIENCE
DETERMINE CUSTOMER PRICE SENSITIVITY
30. With the use of new models
CONTEXT
• The automobile insurance domain works with
cycles presenting stages of profitability & non-
profitability!
• In non-profitability stages, insurance companies
generally have the reflex to increase the premium
in order to reduce their losses!
• Nevertheless, very large increases can have as a
consequence to massively repulse customers
towards the competition. A too high of an attrition
rate could have a negative effect on long-term
profitability of the company!
• A sound management of rate increases is thus of
paramount importance for an insurance company!
• New tools allow the
simulation of insurance
portfolio owned by an
insurer as a function of
the change in the rate
proposed to each of the
insured!
INNOVATION
The availability of such
a tool could result in
increased profits and
would allow to
anticipate attrition
RESULTS
23!Koïos Intelligence Inc.!
CUSTOMER EXPERIENCE
DETERMINE CUSTOMER PRICE SENSITIVITY
31. SEGMENTATION
Client’s data are composed of
20+ variables (individuals,
social, demographics, coverage,
etc.). A segmentation using
unsupervised learning is
necessary.
COMPETITION
METRICS
An important factor in retention
are the prices of competitors. A
proxy indicator for these
external factors shall be
designed and incorporated in
the model
The model must predict, for a
specific profile and external
competition variables, the
probability of retention of a
client
PREDICTION
Extract data using statistical learning
24!Koïos Intelligence Inc.!
CUSTOMER EXPERIENCE
ENSURE CUSTOMER RETENTION
32. Automated learning models stand out from traditional models by their robustness, i.e. their constant reliability in spite more or less important modifications in the data.
The reliability of automated learning models to predict retention has proven effective
Avantages! Disavantages!
Random Forest! Factor selection and
robustness!
Poor prediction
performance!
Logistic Regression! High performance in
prediction!
Absence of robustness!
Gaussian Process! High robustness and
performance of prediction!
Computationally
intense!
Less traditional learning models perform better, generally, than
logistic regression. This is explained notably by the more
important generalization of automated learning models, with the
trade off being the needs for enormous computational
capabilities
Traditional models are generally effective in the case of small changes in the rate of retention, but lack precision in case of more unusual changes. Our
team has worked with models capable of good and robust performance in situation of most extreme rate changes
Learning model to predict retention
25
Koïos Intelligence Inc.!
CUSTOMER EXPERIENCE
ENSURE CUSTOMER RETENTION
33. NOUVEAUTÉS DANS L’UNIVERS DE L’ASSURANCE
ROBOTISATION DES PROCESSUS
TARIFICATION PERSONNALISÉE
GESTION DES RISQUES
OUR TEAM
NEW DEVELOPMENTS IN THE INSURANCE INDUSTRY
PROCESS AUTOMATION
PERSONALIZED PRICING
RISK MANAGEMENT
CUSTOMER EXPERIENCE
6PART
34. OUR TEAM
MOHAMED HANINI
CEO & FOUNDER
MANUEL MORALES
PARTNER & CO-FOUNDER
Mohamed serves North American and global companies in the financial
services. His areas of expertise include risk management, quantitative
trading, business strategy and digital transformations. Mohamed has
done over 10 years of research and development in statistics, finance
and operations research. He taught for 7 years at the University of
Montreal several actuarial and risk management courses such as life
insurance, property and casualty insurance, statistics, financial
mathematics and risk management.
As a Professor in the Department of Mathematics and Statistics of the
University of Montreal, he has accumulated over fifteen years of
experience in collaborative research projects in partnership with key
industry players. He has experience leading collaborative technology
transfer projects in algorithmic trading, market micro-structure modeling,
energy markets and automation in quantitative finance through machine
learning.
Our team is composed of experienced scientists and seasoned consultants in the fields
of mathematics, statistics, economics, computer science, operations research,
quantitative finance and risk management dedicated to innovate through artificial
intelligence technology and to develop business analytics solutions and automation for
today's businesses.
!
!
35. C O N T A C T U S
ADDRESS
5155 Chemin de la rampe
Suite J1249, Montréal (Québec), H3T 2B1
E-MAIL
mohamed.hanini@koiosintelligence.ca
morales@koiosintelligence.ca
514.927.6739
514.730.4975