steps towards the ideal reserving process - read-only

07/05/2021

1

Steps towards the ideal reserving process

Charlie Stone, LCPPhilippa King, AgeasTom Durkin, LCP

25 May 2021

Agenda

• The ideal reserving process

• Opportunities for machine learning to level-up reserving

• Case study: Reserving diagnostics

• Case study: Automated first cut results

• Managing the impact of process changes on the reserving team

25 May 2021 2

1

2

07/05/2021

2

Poll 1

What are your top 3 reserving objectives?

• Earlier identification of trends

• Avoiding reserving surprises

• More time for value added analysis

• Quicker results

• Reduce mundane work

• Improve management information

• Better understanding of reserves

• Reduce expenses

25 May 2021 3

UK firms’ top reserving objectives

25 May 2021 4

Earlier identification

of trends

Avoiding reserving surprises

More time_for value added

analysis

Quickerresults

What areyour objectives?

Reduce expenses

Reduce mundane work

Improve management information

Better understanding

of reserves

We asked which of the following are top priority.

69% 53%

58%

3

4

07/05/2021

3

The future of reserving

• Slicing and dicing

• Earning warning flags

• Scenarios and sensitivities

• Stronger links with the business

• More time to focus on the story

25 May 2021 5

What does the ideal reserving process look like?

Single reliable data

source

Automatic 1st cut

results and diagnostics

Specific external

data

Stay with 1st cut

Specific human input

Manual override

Pri

ori

tisatio

n e

ng

ine

The ideal reserving process

25 May 2021 6

Reserving calculations:

Data Process Analysis

• Majority satisfied with traditional methods

• Easier sensitivity testing and what-if scenarios

• Ability to drill down (into results as well as data)

• Automation to give focus on value add

• Faster process, but few targeting “real time”

• Early warning flags for trends and anomalies

• Single reliable source

• Ability to “drill down” or “slice and dice”

• Enriched with external data

5

6

07/05/2021

4

The ideal reserving process

25 May 2021 7

Wider reserving process:

Communication Feedback loops Governance

• Most firms satisfied governance working well

• Majority of challenge by sub-committees

• Board focus is then on what numbers mean

• Quicker identification of drivers of performance

• Reserving linking to other business functions, incl. pricing, claims, strategy and risk

• Increased focus on story behind the numbers

• Use of scenarios to explain uncertainty

• Some firms targeting a reserving dashboard

Levelling up…

25 May 2021 8

7

8

07/05/2021

5

Setting your “levelling-up” path for reserving

25 May 2021 9

DATA

Reliable, quick triangulated dataIn

creasi

ng s

ophis

ticatio

n

Claim by claim data and policy data

Reliable claim codes and descriptions

Enriched with external data

TRENDS

Automated triangle diagnostics

Individual claims diagnostics

Case reserving early warning system

Monitoring external changes

CALCULATIONS

Streamlined calculations

Assisted assumption selection

Automated 1stcut results

Prioritisationengine

How have you used Machine Learning in your reserving?

• Trend identification

• Assisted assumption selection

• Automated first cut results

• Identifying optimal reserving segmentation

• Non-traditional ML based reserving methods

• Natural language processing of claims forms and notes

• Assessing reserve uncertainty

• Not used at all

10

Poll 2

9

10

07/05/2021

6

Case study 1: Reserving diagnostics

25 May 2021 11

Traditional

• Manual review of triangles

• Time consuming (ie, hours) to review all triangles, so typically consider a selected sample

• Potential to miss features

Levelled-up

• Automated approach, which prioritises top triangles to review

• Quick (ie, minutes), scalable and ability to drill down

• More time to understand the “why?”

What are we looking for?“Fanning out”

25 May 2021 12

“Sticking out”

11

12

07/05/2021

7

What are we looking for?“Fanning out”

25 May 2021 13

Fan direction

Consistency

Number of cohorts

in fan

Drift ratio

Feature engineering:

Using domainknowledge of the data

to create featuresthat make machine learning algorithms

work

Integrating ML effectively

14

Prioritise metricsto review using trend identification (human overlay)

ML highlighting trends for easier identification (providing insight)

Slice and dice and drilldown to understand the drivers of trends (flexibility)

25 May 2021

13

14

07/05/2021

8

A systematic approach to diagnostics

25 May 2021 15

Monitoring experience

Revising assumptions

Reviewing results

• Incurred Loss Ratio

• Paid to Incurred

• Incurred as % dev yr 1

• Incremental Paid to Outstanding

• Paid to Ultimate

• Triangle of Ultimates as % of dev yr 1

• Incurred to Ultimate

• % Non Zero Reported

• Settled Rate

• Incurred as % of previous Yr Ultimate

Each step of reserving process can benefit from systematic approach to

triangle based diagnostics

Poll 3

25 May 2021 16

Which of the following triangle diagnostics do you use in your reserving process?

• Incurred Loss Ratio

• Paid to Incurred

• Claim Frequency

• Incurred as % of dev yr 1

• Incremental Paid to Outstanding

• Average Case Reserve

• Triangle of Ultimates as % of yr 1

• Incurred to Ultimate

• % Non Zero Reported

• Settled Rate

• Incurred as % of previous exercise Ultimate

15

16

07/05/2021

9

Levelled-up

• A new automated reserving engine, prepares reserves in less than 5 minutes

• Engine scores the quality of each reserve estimate for a range of methods and assumptions

• Scores used to prioritise classes to review

Case study 2: Automated first cut results

25 May 2021 17

Traditional

• Manual selection of reserving assumptions and methods

• Time consuming (ie, hours) to apply multiple methods well for a large number of classes

• Little quantitative information to justify one assumption or method over another

ML sets key assumptions – IEULR Example

18

Piecewise linear regression with automated changepoints

+Quarterly seasonality

+

Accident quarter

2010 Q1 2015 Q1 2020 Q1

25 May 2021

Forecast 2020 IEULRs

17

18

07/05/2021

10

ML sets key assumptions – IEULR Example

19

Benefits

• Consistent, objective and easily scalable.

• More accurate predictions.

• Can be run alongside the existing process.

• Used to identify and challenge where assumptions are inconsistent with purely data driven view.

• Improved understanding of volatility in ULRs, feed into UW Risk.

• Other applications, eg as part of claims diagnostics monitoring.

Accident quarter

2010 Q1 2015 Q1 2020 Q1

25 May 2021

Forecast 2020 IEULRs

20

Scoring the quality of reserve estimates

25 May 2021

Volatility

Options include:

• Rollforward volatility

– Movement in ultimates between analyses

• Sub sample volatility

– Repeat the reserving process on samples from the full set of individual claims data

Method and assumption set is “high quality” where it has low volatility and low bias

Bias

Options include:

• ‘Recent experience will continue’

– Emerging experience error

• ‘Nothing has changed’

– Back-testing error

• ‘Wisdom of crowds’

– Difference from other segmentations

Increasing bias

Incre

asi

ng

vo

latil

ity

Incurred DFM: All yr WA

19

20

07/05/2021

11

Managing the impact of reserving process changes

25 May 2021 21

Quality

CostTime

Fast + High Quality-> Expensive

Cheap + High Quality-> Slow/Low Priority

Cheap + Fast -> Low Quality

The Reserving Triangle

25 May 2021 22

Insight

DevelopmentEfficiency

21

22

07/05/2021

12


25 May 2021 23

Insight


• Explain drivers of

experience

• Quantify uncertainty

• Spot trends early

• Get the story right


25 May 2021 24

Insight


• High attrition

• Succession challenges

• Missed opportunities to

add value

23

24

07/05/2021

13


25 May 2021 25

Insight


Efficient + Insightful ->Reliance on senior actuaries

Insightful + Developing ->Takes time

Efficient + Developing =Can’t look beyond existing process

Managing reserving process changes

25 May 2021 26

Insight


Levelling-up the reserving process using ML to provide insight more efficiently:• Combined human-machine approach• Needs effective controls

Efficient & insightful reserving produced

by a skilled & resilient team.

Focus of reserving process improvement in recent years. Training centred on running the process.

Levelling-up approach gives time to upskill team to shift from

running a process to providing insight to the wider business.

25

26

07/05/2021

14

Any questions?

• The ideal reserving process

• Opportunities for machine learning to level-up reserving

• Case study: Reserving diagnostics

• Case study: Automatic first cut results

• Managing the impact of changes in the reserving process

25 May 2021 27

25 May 2021 28

The views expressed in this presentation are those of invited contributors and not necessarily those of the IFoA. The IFoA do not endorse any of the views stated, nor any claims or representations made in this presentation and accept no responsibility or liability to any person for loss or damage suffered as a consequence of their placing reliance upon any view, claim or representation made in this presentation.

The information and expressions of opinion contained in this publication are not intended to be a comprehensive study, nor to provide actuarial advice or advice of any nature and should not be treated as a substitute for specific advice concerning individual situations. On no account may any part of this presentation be reproduced without the written permission of the authors.

Questions Comments

27

28

steps towards the ideal reserving process - read-only

Documents