machine learning: challenges and opportunities for non-life ......2019/06/26 · reacfin breakfast...

Reacfin Breakfast - Machine learning: Challenges and opportunities for non-life pricing and underwriting

IMPORTANT : This presentation is only the supporting document of an oral presentation. It is not intended to be exhaustive. Quoting or using this document on its own might bemisleading. Furthermore, although the authors have been careful in the selection of their sources and assumptions, the authors cannot guarantee that all information inthe document are exact or correct. As a result, these materials may not be used by anybody except their authors nor should they be relied upon in any way for any purposeother than as contemplated by a written agreement with Reacfin.

Brussels - 26th June 2019

Machine learning: Challenges and opportunities for non-life pricing and underwriting

Please read the important disclaimer at the end of this presentation Strictly Confidential

By Samuel Mahy [email protected] Maréchal [email protected]

mailto:[email protected]



1. Introduction and context

2. How to create competitive advantages with Machine Learning for insurance companies?

3. How to solve the Machine Learning challenges for companies?

4. Using machine learning in pricing and underwritting: how to start?

5. Conclusions

Topics to be covered today

P. 2



a. What are the Challenges in non-life insurance?

b. What’s machine learning ?

c. Where is Machine Learning used in insurance?




5. Conclusions


P. 3


To adress these challenges, Insurers have to

Innovate in product development and surrounding services

Capture and identify relevant featuresfor pricing models

Adapt faster to market changes (identification, building of new models, faster deployement)

Optimise retention and renewalpricing

Non-life insurers are facing many challenges putting pressure on their business model and profitability

P. 4

Increasing competition

Commoditisation of insurance products

Sophistication in pricing

Pricing comparison systems

Availability of new data sources

External data (IoT, open data,…)

Use of unstructured data

New customers behavior

Digitalisation of underwriting process

Direct vs Brokers

Focus on price (made possible

thanks to pricing comparison systems)

Main challenges faces by insurance companies


In order to face these challenges, key points of differentiation must be developped by insurers

P. 5

Competitive advantagesin the future

Creative sourcing of data (new sources of external data, behavior-

influencing data monitoring)

Creative usage of data

Distinctiveness of analytic methods

Advanced analytics : far beyond traditionalactuarial sciences

Natural langage processing,

Image processing, …

Large data storage and management

Technology changes muchfaster than people

Insurers should not only invest in analytics technologies

Key for insurers is to train & motivate their highly skilled

experts to adopt the newest tools

Make sure people use Advanced Analytics with creativity,

confidence and consistency


Supervised learning: • Inputs and examples of their desired outputs are provided

• The goal is to learn a general rule that maps inputs to outputs.

Given a set of training examples (x1, x2,…, xn, y), where y is the variable to be predicted , what is the most efficient algorithm to best approximate the realizations of y

• 2 main techniques Classification : inputs are divided into two or more classes, and the learner must produce a model that assigns

unseen inputs to one (or multi-label classification) or more of these classes.

Regression: the outputs are continuous rather than discrete.

Unsupervised learning: • No labels are given to the learning algorithm

• The goal is to find structure in its input (discovering hidden patterns in data).

• Main technique Clustering: a set of inputs is to be divided into groups. Unlike in classification, the groups may not be known

beforehand.

What is machine learning?

P. 6

Objectives of Machine Learning (“ML”)

ML algorithms aim at finding by themselves the method that best predicts the outcome of the studied phenomenon.

Supervised vs. Unsupervised learning


Start by assuming the explanatory model is known and key explaining variables are identified.

Objective : confirm the model assumption and calibrate as accurately as possible the model parameters so that errors can be minimized.

Comparing traditional statistical inference and ML approaches

P. 7

Conceptual difference

Statistical inference

techniques

Machine Learning

techniques

Start from lesser assumption

Objective : the algorithm itself identifies the key explanatory variables and their impact on the response variable.

Starting point & objective Implementation approach

Infer the process by which data you have was generated.

Estimate the model parameters which describe the relationship between the explanatory variables and the dependent variable.

You want to know how you can predict what future data will look like w.r.t. some variable.

The approach is to find a function f(x) – an algorithm that operates on x to predict the responses y


Machine Learning and AI is the continuation of the evolution of tools and technologies used by actuaries and statisticians to analyze historical claims data: trying to improve the predictive power of models, solving the same problems with new methods, data and computer power available

Methods used in non-life pricing are evolving at a fast pace

P. 8


MarketingCustomer

ManagementUW Pricing Risk

Pricing Optimisation

Claims

Machine Learning techniques can be applied all along the value chain in insurance and not only for pricing

P. 9

Web-scraping and campaign steering

Customer segmentation (cross sell, up-sell, customer value)

Brokers’ performance evaluation Feature

engineering, Features selection, geodemographic segmentation,

Segmented & targeted Price increase, churn & new business

Retention Segmentation and management

Modeling risk, tariffs, control leakages, simulate impact of tariff changes

Competitor prices, reverse engineering, portfolio monitoring

Fraud detection

New product targets Simplification

quoting process




a. How can we complete the Data Analytics Toolbox with Machine Learning techniques?

i. Defining model error and managing overfitting

ii. Regressing trees

iii.Random forest and boosting

iv.Artificial neural networks

b. How Machine Learning techniques make possible to boost size and type of data sources?

c. How Machine Learning techniques make possible to boost creativity in using data?



5. Conclusions


P. 10


• 𝑌 = 𝛽0 + 𝛽1. 𝑋1+⋯+ 𝛽𝑛. 𝑋𝑛 + 𝜀

• Y is a direct linear combination of explanatory variables

• The errors are assumed to be Normally distributed: 𝜀 ∼ 𝑁 0, 𝜎2

• And so, 𝑌 ∼ 𝑁 𝜇, 𝜎2

• 𝑌 = 𝑔−1 𝛽0 + 𝛽1. 𝑋1+⋯+ 𝛽𝑛. 𝑋𝑛 + 𝜀

• Y is now a function (g-1) of a linear combination of the explanatory variables

• The distribution of the response variable does not need to be Gaussian anymore

It has to be a member of exponential family

• So, we’ll have for instance, 𝑌 ∼ Poi 𝜇 where 𝜇 = exp(𝛽𝑇𝑋)

Generalized Linear Models are still widely used by insurance companies for non-lifepricing and other applications

Distributions

𝐵𝑖𝑛 1, 𝜇

𝑃𝑜𝑖 𝜇

𝑁𝑜𝑟 𝜇, 𝜎2

𝐺𝑎𝑚 𝜇, 𝛼

𝐼𝐺𝑎𝑢 𝜇, 𝜎2

Linear Model (“LM”)

Generalized Linear Model (“GLM”)

P. 11


ML method minimize an error (or loss) function whereas statistical methods maximize likelihood

In ML, the error function which is minimized depends on the context.

In most cases, we can simply choose the sum of squared errors:

Error ො𝑦 =

𝑖=1

𝑛

ෝ𝑦𝑖 − 𝑦𝑖2 ,

where 𝑦𝑖 is the 𝑖th observation and ෝ𝑦𝑖 is the corresponding prediction.

However, for insurance applications, we must carefully choose our error function. E.g. when wewant to predict Poisson frequencies, it is better to instead consider the Poisson deviance statistics:

Error መ𝜆 =

𝑖=1

𝑛

𝑁𝑖 log𝑁𝑖መ𝜆𝑖 𝑣𝑖

− (𝑁𝑖 − መ𝜆𝑖 𝑣𝑖) ,

where 𝑁𝑖is the 𝑖th observation and 𝑣𝑖 and መ𝜆𝑖 are the corresponding exposure and predicted frequency.

Machine learning technique’s main focus is on prediction and therefore aim at minimizing an error function

P. 12

Which loss function to choose?


• When modelling, we should be sensibilized with overfitting/lack of parcimony.

• It occurs when a statistical model describes random error or noise instead of the underlying relationship.

• The goodness-of-fit indicators show a good result on the dataset used for the model calibration, but the predictive power is bad.

• Example: when trying to explain data variability using a set of explanatory variables, the more variables you use, the better are the 𝑅2, the residual sum of square, etc.

• One way to deal with this issue is to define goodness-of-fit indicators which take into account the number of parameters of the model and apply penalization, such the Akaike and Bayesian information criteria

• But these solutions are not satisfying.

• The choice of a penalization function is arbitrary! Why should it take these forms?

Overfitting deteriorates the predictive power of the models…

P. 13

The overfitting problem

Goodness-of-fit indicators


Use two different datasets:

• A training set to calibrate the model,

• A test set (or validation set) to assess the model’s predictive ability.

…which can be improved by separating the data into a training set and a test set

P. 14

Two different kinds of errors are defined:• The training error is calculated by applying

the model to the observations used in its calibration

• The test error is the average error that results from using the model to predict the response on a new observation, one that was not used in calibrating the model.

The training error decreases with model complexity whereas the test error tends to increase when the level of model complexity creates overfitting

NB: a better solution consists in using cross-validation

A better solution


Tree enables to segment the predictor spaceinto a number of simple regions definedaccording to the covariates

Splitting rules can be summarized in a treeview

For each region the prediction is set as the region average

The root node in orange:

• at the top of the tree

• contains the whole population

The splitting rules set aim at segmenting the predictor space into a number of simple regions.

The leaves nodes in green at the bottom of the tree: that is a node that is not further split.

A first simple ML model: Classification and regression trees

Splitting rulesPurpose

Definitions

P. 15


Regression Tree Algorithm

Define a loss/error (or objective) function and

Try to find regions 𝑅1, 𝑅2, … , 𝑅𝐽 that minimize (or maximize) the function retained

All possible regions definitions can of course not be considered The tree algorithm therefore :

• Starts with the global population and find the optimal split of the predictor at that level using the entire population

• The same process is then applied on each sub-population

If we use the residual sum of square as loss function:

𝐸 𝑇 = 𝑆𝑆𝑇 =

𝑖∈𝑁𝑜𝑑𝑒 𝑇

𝒚𝑖 − ഥ𝒚𝑇2

𝐸 𝑅1 = 𝑆𝑆𝑅1 =

𝑖∈𝑁𝑜𝑑𝑒 𝑅1

𝒚𝑖 − ഥ𝒚12 𝐸 𝑅2 = 𝑆𝑆𝑅2 =

𝑖∈𝑁𝑜𝑑𝑒 𝑅2

𝒚𝑖 − ഥ𝒚22

The optimal splitting variable and point are then obtained through the maximisation of:

𝛥𝐼 = 𝑆𝑆𝑇 − (𝑆𝑆𝑅1 + 𝑆𝑆𝑅2)

Main idea

Optimal splitting

P. 16


Bootstrap aggregation, or Bagging, is a general-purpose procedure for reducing the variance of a statistical learning method

Frequently used in the context of decision trees. Recall that given a set of n independent

observations 𝑍1, 𝑍2, … , 𝑍𝑛 each with variance 𝜎2, the variance of the mean ҧ𝑍 of the observations is

given by 𝜎2

𝑛.

Averaging a set of observations reduces variance. Usually multiple training sets are not at disposal

1. Bootstrap, by taking repeated samples from the (single) training data set.

2. Generate B different training data sets. 3. Train our method on the 𝑏th bootstrapped

training set in order to get መ𝑓𝑏 𝑥 the predictionat point x.

4. We then average all the predictions to obtain :

መ𝑓𝑏𝑎𝑔 𝑥 =1

𝐵

𝑏=1

𝐵

መ𝑓𝑏 𝑥

Bootstrap aggregation (Bagging) allows for variance reduction by averaging over severalregression trees

P. 17

Algorithm

Main idea

…….Bootstrap 1 Bootstrap 2 Bootstrap B

…..1 3 m2Training set

….1 n2 3 ….1 n2 3 ….1 n2 3

Draw nwith replacement

….…መ𝑓1𝑓 መ𝑓2𝑓 መ𝑓𝐵𝑓

መ𝑓𝑏𝑎𝑔

Bootstrap sets

Predictions

Averaging allthe predictions


1. Set መ𝑓 𝑥 = 0 and 𝑟𝑖 = 𝑦𝑖 for all 𝑖 in the training set

2. For 𝑏 = 1, 2, 3, … , 𝐵, repeat :

• Fit a tree መ𝑓𝑏 with 𝑑 splits (𝑑 + 1 terminal nodes) to the training data 𝑋, 𝑟

• Update መ𝑓 by adding in a reduced (shrunken) version of the new tree:

መ𝑓 𝑥 ← መ𝑓 𝑥 + 𝜆 መ𝑓𝑏 𝑥

• Update the residuals:

𝑟𝑖 ← 𝑟𝑖 − 𝜆 መ𝑓𝑏 𝑥𝑖

3. The final model is provided by

መ𝑓 𝑥 =

𝑏=1

𝐵

𝜆 መ𝑓𝑏 𝑥

Boosting allows to learn slowly by fitting rather small decision trees to the residuals from the model

P. 18

Algorithm

…..1 3 m2Training set

….…መ𝑓1𝑓 መ𝑓2𝑓 መ𝑓𝐵𝑓

መ𝑓𝐵𝑜𝑜𝑠𝑡

Predictions on residuals

𝑟1 𝑟2 ….… 𝑟𝐵Update

residuals

Summing part ofthe predictions


What are Neural Networks (NN)?

• They are a method of programming computers, as Random Forest, Support Vector Machine…

• They are often used to perform pattern recognition (unsupervised learning)

• NN can learn on their own and adapt to changing conditions (based on the data)

• NN are inspired by the biological nervous systems such as human brain’s information processing mechanism : they are composed of a large number of interconnected processing elements (neurons) working together to solve problems.

Artificial neural networks expands the perspective of ML and can be used for supervised and unsupervised learning

P. 19

Neuron

X1

X2

Xn

INPUTS OUTPUT

W1

W2

Wn

Description of a neuron

• An artificial neuron is an element with several inputs and one output

• The neuron has two modes of operation:

Training mode (calibration) :

– neuron can be trained to fire (or not) depending on the inputs

– e.g. pattern recognition : associate outputs with input patterns

Using mode (prediction) :

– e.g. pattern recognition : identify input pattern and try to output the associated output pattern


Network representation : connected neurons with different layers

• Input layer (left) containing the input neurons

• Hidden layer(s) (middle)

Neurons in this layer are neither inputs nor outputs this is the origin of the term “Hidden”

Number of layers/neurons :

– In practice, number of layers/neurons determined by trial and error

– One hidden layer is sufficient for most of the problems

– Additional layers can be added if it increases the performance (networks with 2 or more layers are called deep Neural Networks)

• Output layer (right) containing the output neurons

Architecture of Neural Networks

P. 20

Input layer 2 hidden layers Output layer













5. Conclusions


P. 21


Data is present all along the insurance value chain

P. 22

Product

Clients

Financial markets

Claims

Legal

Control

E.g. Use of data to understand the market : use web

information to study competition, market share and compare offers

Data

Performance & continuity

Insurancecompany

E.g Use of data to target clients:

segment needs in function of their characteristics (location, behaviour,

etc.) to propose a relavant insurance offer

E.g Use of financial data to predict

asset values: asset historical data used to calibrate tool and predict asset

market values for optimizing allocation purpose

E.g. Use of external legal data:

automatic follow-up of regulation updates and regulatory trends

E.g.Use of data to control internal

data quality: such a type of control also needs to be put in place for a

regular control

E.g. Use of of data to improve actuarial practices:

measure frequency and claims amount to perform actuarial studies and improve pricing and reserving

calculations


Numerous sources of internal or external data

Data type is different from one content to another

Different categories of data and different sources and types of information

P. 23

Internal data External data

PDF files

Commercial data

Websites

Purchased databases

Mobile dataEmails

Open data

CRMModel calculations

Social media

Data warehouse

Structured data

Unstructured data

Word files Sensor data

Number Text Audio Image/video Other

Structured data : organized and well characterized data that are easy to use because they are well identified.

• E.g. insurer’s policies and claims data

Unstructured data: non-organized data not easy to manipulate and which require much preparation (everything else).

80%

20%


It is key for insurance companies to broaden the data collection perspectives to boost the number of data sources available

P. 24

Machine learning techniques allow to deal with very large amount of data and therefore create opportunities for insurance company to increase the number of features to be analyzed/used in the pricing and underwriting process

Additional data can be obtained through many different sources :

1. Scraping/parsing techniques:

Extract information

automatically from

websites

2. Open data files:

Structured datasets

available to everyone

3. IoT sensor and API technologies:

Connected objects and application

programming interface

4. External data provider

Ready to use data set for

sale

5. Look twice into your own

unstructured data:

Reveal hidden

information from core

(unused) data


To be put in perspective with legal environment and regulation

Opportunities and threats of creative data sourcing for insurance pricing

P. 25

Retail Business

Important segmentation usually already in place

Limited potential for further segmentation

Corporate Business

Only few segmentation variables available

Greater pricing refinement potential

GDPR is in place since2018

Segmentation criteriamust be disclosed

Anti-discrimination

Insurers are limited in their use of data but opportunities exist to expand it













5. Conclusions


P. 26


Once additional data has been collected, new methods and algorithm allow to get the most out of it. Among others:

How to enhance data

P. 27

1. Statistics, ML and feature engineering:

Create structured dataset using initial datasets or charts to

understand data

2. Text mining and NLP

Process of examining large collection of written

resources and methods to perform linguistic

analysis

3. Image processing

Techniques to perform operations on images to enhance its content or

extract information


External sources and new types of data

Data Manipulation

GLM Modeling

Deployment

DB 2

Data Extraction

DB 4

DB 1

DB 3

HOW to introduce external and new types of data in the pricing process?

Usual Non-Life Pricing Process 1. Starts with data extractions2. Followed by some data checks and formatting steps3. Generalised Linear Model adjustement (eventually with Forward or

Backward procedure but not always…)Should we adjust this process because of new and external data?

P. 28


Data Scientist and Actuaries view

P. 29

DB 4

DB 3

DB 2 Data Extraction & MERGE

DB 1

A B C

1 … …

3 … …

5 … …

6 … …

DB1

A D E F

5 … … …

1 … … …

6 … … …

2 … … …

4 … … …

8 … … …

DB2

A B C D E F

1 … … … … …

3 … … … … …

5 … … … … …

6 … … … … …

DB1 DB2

N rows

P variables

Enrich the existing database with new attributes/variables

When External an new data are used in order to enrich the existing database with new attributes/variables :

• External databases should be merged with internal• It requires an adequate merging key

As number of attributes/variables increases :• Overfitting should be managed (Cross-Validation, Regularization)• Features Selection & Engineering even more important


Data Scientist and Actuaries view

P. 30

FeatureSelection

FeatureEngineering

Data Manipulation

Modeling withCV &

Regularization

Statistical or ML methods

Model Evaluation

& Comparis

on

Deployment

DB 4

DB 3

DB 2 Data Extraction & MERGE

DB 1

Additional Steps should appear in the Pricing process (like in machine learning approach)

Features Engineering/Selection: set of methodologies to extract meaningful attributes or features from the raw data

Regularization : Algorithms which are designed to manage database with large number of attributes/variables controlling overfitting


Underwriter and marketing teams

P. 31

FeatureSelection

FeatureEngineering

Data Manipulation

Modeling withCV &

Regularization

Model Evaluation

& Comparis

on

Deployment

DB 1

FeatureSelection

FeatureEngineering

Data Manipulation

Modeling withCV &

RegularizationDB 2

Simplifying the quoting process ?

Replace existing features with new features obtained from external DB

Compare models to measure the adequacy/performance of the new features

As new features are coming from external data provider, the UW form can be simplified (eg. Quick quote system using only vehicle plate numbers)


Features Engineering & Selection

Features Engineering is absolutely known and agreed to be key to success in applied machine learning.

Features Engineering is a Representation Problem• Machine learning algorithms learn a solution to a problem from sample data.• In this context, feature engineering asks: what is the best representation of the sample data to learn a

solution to your problem?

Frequency

Age

Age Frequency

- -

- -

- -

- -

Concept of Feature Engineering

P. 32


Features Engineering & Selection

The results you achieve are a factor of :• the model you choose, • the data you have available • and the features you prepared.

The better the features that you prepare and choose, the better the results you will achieve

Frequency

Age

Frequency

Age

Frequency

Age^2

Linear (eg. GLM) Non-linear (eg.GAM) New Features (eg Age^2)

Poor Model choice and without featuresengineering

More complex model and without featuresengineering

Simpler Model choice BUT features engineering

Feature Engineering vs Model Complexity

P. 33


Feature Engineering combined with external data set

Intuition : there is a correlation between the claims frequency and the distance from the highway

• Data available in the company : addresses

• Features Engineering : convert house addresses into distance from the highway

Highway only?

• No, all the roads where the speed limitation is above 90km/h

Determine the closest point to the highway in relation to the house.

• We need to know the location of the house on a map

• We need to know the location of the highway on a map

Open Street Maps

• Gives the roads’ longitude and latitude at different points.

Example (1/2) : Property Theft Insurance

P. 34


Feature Engineering combined with external data set

If we want to find the distance from the house, we need the coordinates of the house.

Google Maps Geocoding API

• Geocoding is the process of converting addresses into geographic coordinates

Google API is free but with slow performances:

• 2,500 free requests per day

• 50 requests per second (limitation of speed)

• Enable pay-as-you-go billing to unlock higher quotas: $0.50 USD / 1000 additional requests, up to 100,000 daily.

Find the distance between the house and the first road (above 90km/h).

• We build a loop that checks if there is a road in a growing area (in a radius growing from 0 to 4000m with step of 200m)

Example (2/2) : Property Theft Insurance

P. 35





a. Machine learning results can be difficult to interpret

b. Data Analytics must be fit for purpose


5. Conclusions


P. 36


In the case of regression trees, understanding how the model predicts claims’ cost or frequency values for new

data points is not a problem, as it is very intuitive.

In the case of more complexmethods such as Bagging and

Random forests, even understandinghow the model predicts values for new data points is rather difficult.

Things may be even

worse for GBM

and NN.

Some machine learning techniques are black boxes and interpretation of the resultscan be quite difficult

Understanding the results of ML techniques is not easy

Complexity

Interpretability

P. 37


Quant (Actuaries, data scientist,…)

Understanding the results of ML models is nevertheless key for sound business decision-making as many stakeholders use the results of the models

Machine learning techniques usually improve predictive power but at the expense of a certain loss of interpretability Find trade-off between

Other stakeholders

Not necessarily « quantitative people »

Should neverthelessunderstand and trust results to

take decisions

Predictive power Capacity to understand

the results

Ability to take sound decisions based on the

results

High-end questions

Who will use the results? For what purpose? With which impact?

Able to understand the technical details

Trust its outputs based on cross-validation, error

measures and assesment plots

P. 38


How to integrate Machine learning in pricing process?

First solution: stick to existing (e.g. in production) model and use machine learningtechniques as guiding tools for

• Features selection

• Features engineering

Some Machine Learning methods can produce graphs which enable to understand how important a variable is in the prediction, eg. :

• Random forests

• LASSO (penalized regression)

These graphs can therefore be used as pre-modeling approach to explore the data and decide which features we will select in the model

Feature selection and features engineering

Random Forest importance score

P. 39


How to integrate Machine learning in pricing process?

Machine learning techniques usually improve predictive power but at the expense of a certain loss of interpretability.

Some tools can be used in order to help in the interpretation and understanding of the results

For example with random forest, bagging or boosting trees methods

• Identification of variable importance (see supra)

Partial dependence plots(1 or 2 variables)

Residual plots as a function of a variable

Develop “interpretation tools”

P. 40






a. Profitability analysis

b. Competition Analysis

c. Policyholder’s Behaviour

5. Conclusions


P. 41


Commercial Pricing is a permanent multidimensional optimization process under complex constraints where segmentation plays a crucial role

Portfolio profitability by segment, based on:> Cost of risk (e.g. measured through GLMs

or ML techniques) > Portfolio composition (representativeness

of each segment: total portfolio and recent new business)

Competitor prices by segment and own current rates> Position insurer to be more or less

competitive on certain segments

Customer behavior by segment > Elasticity model help estimate pace at

which rates can be increased by segment> Focus Sales & Marketing to increase

retention of better risks> Building conversion rates model to better

target clients

Segmentation and pricing variables> Greater segmentation for greater risk

selectivity and higher profitability> Monitor concentrations of certain risk

types

Constrainsrates of existingportfolio

Constrainsrates of newproduction

Aligned segments

TE

CH

NIC

AL

PR

IC

IN

GC

OM

PET

I-

TIO

NC

LIEN

T

BE

HA

VIO

R

A

B

C

D

SEG

MEN

-TA

TIO

N

E SCENARIO TESTING AND OPTIMISATION

Impact of different scenarios on strategic indicators and optimization

P. 42


Profitability analysis tool Regression trees techniques can be used to compare Risk Premium and Commercial premium

Thanks to regression trees it is possible to identify the variables that are the most relevant toexplain the differences between the risk premium and the current commercial premium

• It helps in defining the most relevant variables that can, for example, then be included in aprofitability heatmap

A

P. 43


Identifying the segments in which the insurance company is well-positioned with respect to its competitors is an important driver of a dynamic pricing process. E.g. Clustering of segments in function of the ranking of the competitors with regression trees

Competition analysis tool Regression trees can be used to identify positioning on market segments and capture price differences

B

Analyze the price dispersion of the specific company with respect to its competitors of wrt respect to the average market price

Reverse engineering of the pricing (structure) of competitors can be enhanced with ML techniques

P. 44


The goal is to explain the conversion / lapse probabilities with some explanatory variables

A dummy variable identifies the policies that were converted / renewed during the year

Traditionaly Generalized Linear Models are used

– E.g. A logistic regression can be performed on this dummy variable and potential explanatory variables

𝑙𝑛𝜋(𝑥1…𝑥𝑛)

1 − 𝜋(𝑥1…𝑥𝑛)= 𝛽0 + 𝛽1𝑥1 +⋯+ 𝛽𝑛𝑥𝑛

Machine learning technique (e.g. GBM) are more and more often used as they usually improve predictions and allow to find more complex patterns

Client behavior ML techniques can help improve the logistic regression

C

P. 45






a. Profitability analysis

b. Competition Analysis

c. Policyholder’s Behaviour

5. Conclusions


P. 46


Conclusions

P. 47

Pricing environment

Competition willincrease

More data will beavailable

Frequency of tariffreview will increase

Future of technicalprice

Pricing will no longer rely only on GLM but

on a set of algorithms/methods

Insurer should be able to :

• Understand and use these methodsadequately

• Apply, compare and deploy these modelsrapidly

Automation required

Insurer can no longer have a process which

lasts 6 months in orderto deploy a new tariff

Automation of the pricing process is one of the key factors of

success


Reacfin’s support

Efficient solutions aligned with your best interest

• Experienced staff with hands-on knowledge and proven track record, incl.

– Large industry experience

– Training capabilities in Belgium and abroad

– Sound knowledge of the products

• Extensive training material (methodologies, exercises and case studies) through various channels (e-learning, slides, notebooks, open source code)

• Largely networked within the industry

• Working in your best interest as independent consultant

What we do… … and why you should consider it

• Feasibility assessments and defining the solution

• Training on statistical models and machine learning techniques with hands-on exercises in R or Python

• Technical pricing (model development and improvement)

• Support for commercial pricing (competition analysis, conversion models, lapse models,…)

• Implementation in open source (R,Python) and proprietary software (SAS, Emblem,…) software

– Existing tools for pricing, dispersion analysis and competition analysis

Helping you developing best market practices for an affordable price

P. 48


Place de l’Université, 25B-1348 Louvain-la-Neuve (Belgium)

T +32 (0) 10 84 07 50

www.reacfin.com

Samuel Mahy

Director – Head of Non-Life

M +32 498 04 23 [email protected]

Xavier Maréchal

CEO

M +32 497 48 98 [email protected]

Contact details

P. 49

http://www.reacfin.com/




Place de l’Université 25B-1348 Louvain-la-Neuve

www.reacfin.com

Disclaimer:

The recipient of this document should treat all

information as strictly confidential and only in the

context stated below. Information may not be

disclosed to any third party without the prior join-

consent of Reacfin.

Estimates given in this presentation are based on our

current knowledge, they can be based upon our

previous experience within the Undertaking, as well as

taking into account similar projects in the same

context as the Undertaking, either locally, within

majority of the EU countries as well as overseas.

This presentation is only the supporting document of

a verbal presentation. Hence, it is not intended to be

exhaustive. Quoting or using this document on its own

might be misleading. As a result, these materials may

not be used by anybody except their authors nor

should they be relied upon in any way for any purpose

other than as contemplated by joint written

agreement with Reacfin.

machine learning: challenges and opportunities for non-life ......2019/06/26 · reacfin breakfast...

Documents