the journey towards data driven decisions - swiss re5bdfda64-33f1-42a4... · system- and data...
Post on 16-Aug-2020
0 Views
Preview:
TRANSCRIPT
The journey towards data driven decisions – how can we benefit from data (even if we don’t own them)?
Karl Ove Aarbu, Head of Customer & Claim Analytics, Tryg
2
2
•Second largest Nordic non-life player
•DKK 18bn of premiums (EUR 2.41bn or GBP
2.16bn)
•Largest player in Denmark, 4th in Norway and
5th in Sweden
•Approx market cap of DKK 60bn (EUR 8.1bn)
NorwayMarket position: #4
Market share: 13.3%CR in Q2 18: 82.8%
SwedenMarket position: #5Market share: 3.1%CR in Q2 18: 98.3%
DenmarkMarket position: #1Market share: 22%
CR in Q2 18: 77.2%
Business split 2017
Gross premium split by products 2017
Percentage
Tryg at a glance
3 1. september 2019
4 1. september 2019
5
…data must be refined ….build CustomerTimeLineData…act when the customer is responsive
1. september 2019
Call centresystem
Claimsystem
WebdataProduction
systemSales
systemetc M
CustomerTimeLine Data
Right time
Right person
Right channel
Right message
Time
Churn prob.
«LoveYouCall»
6
…and our actions must focus on maximizing CLV…
1. september 2019
𝑴𝒂𝒙𝒊𝒎𝒊𝒛𝒆 𝑪𝑳𝑽 =
𝒊=𝟏
𝑴
{𝑷𝒓𝒆𝒎𝒊𝒖𝒎− 𝑬𝒙𝒑𝒆𝒄𝒕𝒆𝒅(𝑪𝒍𝒂𝒊𝒎𝒄𝒐𝒔𝒕𝒔) − 𝑨𝒅𝒎_𝒄𝒐𝒔𝒕} ∗ 𝑬𝒙𝒑𝒆𝒄𝒕𝒆𝒅(𝑳𝒊𝒇𝒆𝒕𝒊𝒎𝒆)
𝑨𝒅𝒎_𝒄𝒐𝒔𝒕 (decrease adm. costs)STP (Straight Through
Processing)
Elements of CLV Action areas
Expected net saving with manual handling
Manual or automatic
Actions (treatments)AI Models (examples)
M (increase #customers) Find profitable segments in portfolio
Customer value models
Online banners
𝑳𝒊𝒇𝒆𝒕𝒊𝒎𝒆 (Increase lifetime
existing customers)Churn/Survival
Short term churn model
“love you” call
“love you sms/email”
“pleasant surprise”
Facebook “twins”Linkedin “twins”
Target discounts/”pleasant
surprise
7
Mindset/Culture Technology and data Analytics/ML/AI methods -
predictions
Behavioral understanding
and policy optimization
KPIs
Maximize
CustomerLifeTimeValue
System- and data silos.
Hard coded rules.
Customer TimeLine and
Analytics as a Service (cloud)
GLM Models –
static features –
long term
Always full blown
implementation
Experimentation and
Reinforcement
Short term Prescriptive
ML-models
..so to become data driven…we have to climb some stairs…
8
…but how do we know that the «action» works?
1. Predictions are not
Prescriptions
2. AB testing in order
check effectiveness of
actions…
3. ..always relevant to
not act at all…
1. september 2019
A B
x
9
Business problem: High churn..
1. september 2019
Random split (A_B)
Profitable customers
Bring business problem to data=Model
Profitable «disloyal» customers
Control group Love you calls Email/SMS
10
..and it works….
1. september 2019
• Significant differences in
churn rate between
control and treatment
group
• Most effective among
youngest and oldest
customers..
Age category
Churn
rate
90 d
ays
aft
er
treatm
ent
Group
11
..event data are extremely important!
1. Some data are more important than
others
1. Static or determenistic: Age,
gender, adress…..
2. Semi-static: Number of products,
premium…..
3. Events: Webvisit,
outbound/inbound call, claim
event……..
2. CTL should include all types of data –
especially events
1. september 2019
0
1
2
3
4
5
6
7
8
9
Webvisit Not webvisit
Churn rate 30 days after treatment
Treatment Control
2.4 times more churn in web group
12
..and what about using data you do not own?….
1. Seed audience – Facebook / Linkedin
1. september 2019
Profitable customers
Secure login«MyPage»
«Twins»
Banner
13
..and…..we consider to stop this “treatment/campaign”… should
we?
…how large should the control group be?….
..…how should we design the experiment?...
..signs of success – when we naturally asks….
14
…summary….the road to data driven decisions requires….Summary…..
1. Getting your data to sing is no walk in the
park….
1. Mindset and culture change
2. New technologies – analytics as a service!
3. Short term prescription models that aims
towards maximizing CLV
4. Experimentation, Experimentation,
Experimentation
2. Everlasting process – business, technology,
datawarehouse and analytics hand in hand..?
top related