System Challenges in Fixed Index Annuity US GAAP Implementation - - PDF document

Equity-Based Insurance Guarantees Conference

• Nov. 5-6, 2018

Chicago, IL

System Challenges in Fixed Index Annuity US GAAP Implementation Peter Phillips

SOA Antitrust Compliance Guidelines SOA Presentation Disclaimer

Sponsored by

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

System Challenges in Fixed Index Annuity US GAAP Implementation

Peter Phillips Jerry Mao

November 6, 2018 1:30-2:30 p.m.

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

2

Legal Disclaimer

Aon Benfield Securities, Inc. and its affiliates (“Aon”) are not currently registered in any securities advisory capacity in any Canadian jurisdictions. These materials are not being provided under any advisory mandate and shall not be viewed or construed as any type of investment advice, including but not limited to investment advice tailored to you or any other party. Therefore, these materials shall not be relied upon as investment advice in any circumstances. These materials contain information that is confidential or proprietary to Aon and shall not be disclosed to any third party without the express written consent of Aon. This document is the confidential property of Aon Benfield Securities,. (“Aon”), has been prepared by Aon for informational purposes only and is intended only for the designated recipient. As a condition to reviewing this document, the recipient agrees that without the prior written consent of Aon, which may be withheld for any reason, the recipient will not copy the document or any of its contents, and will not disclose or disseminate the document or any of its contents to (i) any third party, or (ii) any person within recipient’s organization who does not have a need to know in connection with the express business purpose for which the document is being provided to recipient. If the recipient is legally compelled to disclose this document or any of its contents, it will promptly give notice to Aon and will reasonably cooperate with Aon in any attempts by Aon to obtain a protective order or otherwise limit disclosure. Upon request by Aon, the recipient will promptly return or destroy the document and any copies it has made with Aon’s consent (as described above), provided that recipient may maintain, in strict confidentiality, such copy or copies as required by law or regulation. Aon makes no representation of any kind as to the suitability of the products or services described in this document for any entity in any jurisdiction. The recipient is advised to undertake its own review of the legal, regulatory, tax, accounting and actuarial implications of the products and services described in this document, as Aon does not provide legal, regulatory, tax, accounting or actuarial opinions. This document should not be considered an offer to sell or a solicitation of any agreement to purchase any security. All securities advice, products or services are offered solely through Aon Benfield Securities, Inc. or an appropriately licensed affiliate.

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

3

Agenda

Section 1 FIA Reporting Changes Section 2 Implementation Challenges Section 3 Implementation Strategies

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

4

Agenda

Section 1 FIA Reporting Changes Section 2 Implementation Challenges Section 3 Implementation Strategies

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

5

Introduction

• Accounting Standards Update 2018-12 outlines 4 areas of improvements for long-duration contracts.
• December 15, 2020 implementation date poses challenges from a system perspective.

DAC Amortization Disclosures Assumptions Market Risk Benefits

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

6

US GAAP Targeted Improvements

• Review and update cash flow measurement assumptions at least annually

– Recognized in net income

• Review and update discount rate assumption at each reporting date

– Recognized in other comprehensive income

• Deferred acquisition costs (DAC) amortization

– Constant level basis over the expected term (simplification)

• Disclosures

– Disaggregated rollforwards

• Reserve, account balances, separate account liabilities, DAC
• Market risk benefits

– Significant inputs, judgements, assumptions, and methods

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

7

FIA US GAAP Targeted Improvements

• Measurement of market risk benefits

The amendments require that an insurance entity measure all market risk benefits associated with deposit (or account balance) contracts at fair value. The portion of any change in fair value attributable to a change in the instrument- specific credit risk is required to be recognized in other comprehensive income.

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

8

Agenda

Section 1 FIA Reporting Changes Section 2 Implementation Challenges Section 3 Implementation Strategies

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

9

Common Challenges

• Computation time

– Large-scale seriatim calculations or Monte Carlo simulations

• Memory and data bottlenecks

– Inefficient use of memory and unnecessary data input / output

• Multitude of special-purpose systems

– Lack of integration and high-level control of various systems and platforms

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

10

Exotic Options in FIA

• Options are embedded in FIA contracts to back crediting rates

– Non-exotic options: call spreads – Exotic options: Asian options, cliquets (forward starting options) – Reported according to ASC 815 (Embedded Derivatives)

• Option valuation methods:

– Closed-form: not always available, numerical methods required – Approximation: not always reliable, further complexity in approximation methodology – Monte Carlo simulation: most general method, slow under legacy systems

End of Year 2011 2012 2013 2014 2015 2016 2017 S&P 500 Return

• 0.003%

13.41% 29.60% 11.39%

• 0.73%

9.54% 19.42% Rate Credited 0% 5% 5% 5% 0% 5% 5%

Sample Illustration (Annual Point-to-Point with 5% Cap)

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

11

Market Risk Benefits

• Most GMxB riders are going to be classified as market risk benefits according to ASC 944 and valued
• n a fair-value basis
• GLWB is an increasingly popular feature for FIA contracts
• Switching to fair value measurement will impact valuation, hedging, and reported results

– Hedging mismatch if currently hedge is on current accounting basis

• Increase balance sheet volatility
• Significant increase in computational load
• Modifications to existing processes

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

12

Existing Production Processes

• The existing accounting processes ASC 815 and ASC 944 are complex enough, projection of financial

statements and hedging are usually nested simulations

• New reporting changes further increase the demand for computation resources

Hedging reporting (ASC 815) Embedded derivatives reporting (ASC 815) Host contract reporting (ASC 944) Additional liability reporting (ASC 944) Real-World Scenarios and Sensitivities Market-consistent Scenarios and Shocks SOP 03-1 Scenarios

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

13

Potential Changes

• The following changes may be required:

– Replace old real-world valuation methodology with the new fair-value methodology – Integrate market-consistent scenarios – Update hedging methodology – Update and consolidate reporting processes – Restatement of financial results

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

14

Change Management

• Impact analysis

– Systems and processes impacted – New processes required – Estimate computational load and run times – Estimate human resource requirements – Assessing the impact on financial results

• Transition planning

– Update existing systems and processes or – Introduce new systems and processes – Running multiple systems or multiple versions in parallel – Pathway from development to production

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

15

What’s in a system?

• To understand how the system is impacted, we need to take a closer look at the system itself
• A modern computation platform can be described as a vertical stack

User Front End Applications Models and Analytics Middle Layer Back End Services Computation Hardware Input File I/O Output

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

16

User

• User is at the top of the stack
• Legacy systems

– Poor user experience (UX) – Steep learning curve – Poor documentation – Low-level language skills needed

• Modern systems

– Intelligently designed UX – Gentle learning curve – Integrated documentation system – Visual programming

What’s in a system?

User Models and Analytics Front End Applications Middle Layer Back End Services Computation Hardware Data Flow Input File I/O Output

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

17

Models and Analytics

• Legacy systems

– Low flexibility – Inadequate built-in reporting and visualization capabilities – Error-prone code by programmers – Brute force design with no concern for computational or memory efficiency – Difficult or impossible to drilldown bedrock or debug

• Modern systems

– Controlled flexibility – Models implemented by industry experts – Optimized for speed and efficiency – Deep analytics and visualization – Easy to drilldown and debug

What’s in a system?

User Models and Analytics Front End Applications Middle Layer Back End Services Computation Hardware Data Flow Input File I/O Output

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

18

Front End Applications

• Applications users interact with
• Legacy systems

– Specialized application with an incomplete set of tools or – General programming environment without any domain-specific optimization

• Modern systems

– General-purpose applications designed for the industry – Complete set of modelling, business intelligence, database, automation tools

What’s in a system?

User Models and Analytics Front End Applications Middle Layer Back End Services Computation Hardware Data Flow Input File I/O Output

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

19

Behind-the-Scene Layers

• Lacking in many legacy systems
• Middle Layer

– Optimization of the complied code – Middleware (distribution of calculations to the grid)

• Back End Services

– Database system – System monitoring, logging, and versioning – Automation, job scheduling and prioritizing – User access rights, OS admin

• Computation Hardware

– GPUs, network, and servers

• Data Flow

– Intra-system data I/O – Interface for external systems

What’s in a system?

User Models and Analytics Front End Applications Middle Layer Back End Services Computation Hardware Data Flow

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

20

Traditional Development Cycle

• Traditional IT-driven model
• Many iterations required to get close to what the business wants
• Prolonged development cycle
• Slow to react to changes

Prepare requirements Business IT/Vendor Review requirement Project plan Review project plan Prepare detailed spec Review spec Implementation User testing Production

Iterations

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

21

Rapid Development Cycle

• New business-driven rapid development model
• Separation of business logic and system functionality
• IT development focuses on system performance and new functionality
• Accelerates development by allowing fast iterations

Prepare requirements Prepare project plan Implementation User testing / documentation Production Business IT/Vendor System development System testing System update

Iterations

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

22

System Concerns during Transition

• Is the system capable of running multiple processes at the same time?
• Does the system provide an easy path from development to production?
• Does the system have integrated logs and controls covering data, models, scripts and reports?

Development Testing Production

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

23

Agenda

Section 1 FIA Reporting Changes Section 2 Implementation Challenges Section 3 Implementation Strategies

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

24

Implementation Strategies

• Calculation efficiency

– High performance computing – The small but difficult things

• Memory efficiency

– Time-space tradeoff – Other factors

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

25

Calculation Efficiency

• Parallelism and GPU (Graphics Processing Unit)
• GPUs are designed to handle massively parallel tasks such as processing millions of

insurance policies or thousands of scenarios

CPU GPU

Several complex cores Thousands of simpler cores

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

26

CPU

• Optimized for serial processing
• Applications

– Operating systems – General computing

GPU

• Built for parallel processing
• Applications

– Monte Carlo simulation – Big data – Machine learning

Calculation Efficiency

CPU GPU

• Parallelism +

+ Control Logic Complexity -

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

27

GPUs

• GPU (Graphical Processing Unit) started in 1999, with the launch of GeForce 256, as an add in card used for 3d rendering in video games,

which has now grown in 2017 to a $109B a year industry with over 600M users*

• Today GPUs are used in fields as diverse as artificial intelligence, machine learning, autonomous vehicles, oil exploration, image

processing, statistics, algebra, 3D reconstruction, medical imaging, finance, etc

• Many processes that previously took days to be completed serially can now can be done in minutes or seconds using GPUs because all

the jobs can be done in parallel.

• CPUs (Central Processing Unit) are optimized for sequential tasks
• GPUs are optimized for compute intensive parallelizable (e.g. Monte Carlo simulation) tasks
• Mythbusters’ Youtube video—GPU vs CPU
• https://www.youtube.com/watch?v=-P28LKWTzrI

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

28 GPU- Accelerat ed Comput ing 107 105 1.1X per year 103 Single-t hreaded performance 1.5X per year 1980 1990 2000 2010 2020

40 Years of CPU Trend Data

Original data up to the year 2010 collected and plotted by M. Horowitz, F. Labonte, O. S hacham, K. Olukotun, L. Hammond, and C. Batten New plot and data collected for 2010-2015 by K. Rupp

370 PF Total GPU FLOPS

• f Top 50

S ystems 15X in 5 Yrs

2013 2018

8, 500 GTC Registrations 4X in 5 Yrs

2013 2018

820,000 GPU Developers 10X in 5 Yrs

2013 2018

8 M CUDA Downloads 5X in 5 Yrs

2013 2018

The Rise of GPU Computing

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

29

The Age of Accelerated Computing Continued…

• Fermi GPU S

erver 2013

• HPC Applicat ions Amber

12

• NAMD

2.9

• GPU
• Acceleration
• St ack
• cuBLAS

5.0

• cuFFT

5.0

• NPP

5.0

• CUDA

5.0

• cuRAND

5.0

• cuSPARSE

5.0

• Res Mgr

R304

• BaseOS

CentOS6.2

• Volt a GPU Server 2018
• HPC Applicat ions Amber

16 CHROMA 2018 Gyrokinet ic TC 2017 LAMMPS 2018 MILC 2018 NAMD 2.13 Quant um Esp 6. 1 SPECFEM3D 2018

• GPU
• Acceleration
• St ack
• cuBLAS

9.0

• cuFFT

9.0

• NPP

9.0

• CUDA

9.0

• cuRAND

9.0

• cuSPARSE

9.0

• Res Mgr

R384

• BaseOS

Ubuntu 16.04

• 100
• GPU–Accelerat ed

Comput ing

• 10
• Mo o r e ’ s Law
• 50%

per year

• CPU
• 1
• 2013 2018
• Measured performance of Amber, CHROMA, GTC, LAMMPS, MILC, NAMD,

Quant um Espresso, SPECFEM3D

In 5 years, memory size went from 6 GB to 516 GB, and memory speed went from 192 GB/s to 900 GB/s, and from throughput from 1.5 Tflops to 16 Tflops in single precision ---eye watering figures

Measured performance of Amber, Chroma, GTC, LAMMPs, MILC, NAMD, Quantum Expresso, SPECFEM3D

• 10% a

y e ar

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

30

A Common Sight at Many Companies…

• TRADITIONAL
• HPC CLUS

TER

600 Dual-CPU S ervers 360 kW

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

31

The More GPUs You Use the More $$$ You Save

• NVIDIA
• TES

LA V100

• BIG S

AVINGS FOR HPC

• 30 Quad-GPU Servers
• 48 kW
• 1/ 5 the Cost
• 1/ 7 the S

pace

• 1/ 7 the Power Consumption

It’s Green Technology

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

32

Something is Going On Here…

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

33

Calculation Efficiency

• To realize the full potential of GPUs:

– Placement strategy is needed to distribute calculations to multiple CPU and GPU devices

• Strategy is highly dependent on the nature of the calculations
• Would this calculation run faster on CPUs or GPUs?
• How much memory is required for this calculation? How fast is the network?
• Algorithm still matters but often hidden

– GPUs need to be saturated

• Have enough parallel calculations to utilize all GPU cores

– Fault tolerance

• Results cannot be affected by hardware errors or failures

– Scalability

• All these issues need to be addressed by the middleware that sits between the user’s logic, the network,

data, GPUs and storage devices.

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

34

Calculation Efficiency

• Precision/speed tradeoff
• There is no silver bullet single correct answer
• Greeks calculations for GMxB

– Finite difference estimator: [f(x + ε) – f(x)] / ε – Determination of ε is not easy – Cancellation due to lack of precision can be a serious problem – Other approaches: auto diff, complex number, dual number

• Rigorous testing is really necessary

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

35

Memory Efficiency

• Space/time tradeoff

– An important tradeoff in computer science

• Classic examples:

– Disk I/O with or without compression – Table lookup vs recalculation

• The optimal solution depends on the relative cost/availability of memory and the nature of

computational power

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

36

• Automatic differentiation is a technique to differentiate a function in the form of a computer program
• Widely used in machine learning and computational finance (pathwise differentiation)
• More stable than finite difference
• Two different modes: forward and reverse

function f(x_1, x_2): x_3 = 2 * x_1 x_4 = x_2^2 x_5 = x_3 + x_4 return x_5

Example: Automatic Differentiation

x_5 x_3 x_4

+

x_1 2 x_2 x_2

× ×

f(x1, x2) = 2x1 + x2

2

Computation Graph

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

37

• Forward mode starts from the inputs
• Differentiate with respect to x1

Example: Automatic Differentiation

x_5 x_3 x_4

+

x_1 2 x_2 x_2

× ×

dx_1/dx_1 = 1 dx_3/dx_1 = 2 * dx_1/dx_1 = 2 2 * x_2 * dx_2 = df/dx_1 = dx_3/dx_1 + dx_4/dx_1 = 2

1.0 0.5 0.5

Forward sweep (x_1)

function f(x_1, x_2): x_3 = 2 * x_1 x_4 = x_2^2 x_5 = x_3 + x_4 return x_5

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

38

• Differentiate with respect to x2
• One forward sweep for each partial derivative

Example: Automatic Differentiation

x_5 x_3 x_4

+

x_1 2 x_2 x_2

× ×

1 1 dx_4/dx_2 = 2*x_2*dx_2 = 1 df/dx_2 = dx_3/dx_2 + dx_4/dx_2 = 1

0.5 0.5 1.0

Forward sweep (x_2)

function f(x_1, x_2): x_3 = 2 * x_1 x_4 = x_2^2 x_5 = x_3 + x_4 return x_5

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

39

• Backward mode consists of two steps: memorization and backward sweep
• Memorization step stores all the interim calculations in memory i.e. building the Wengert list
• Checkpointing strategies can be used to reduce the memory footprint

Example: Automatic Differentiation

x_5:

x_3+x_4

x_3:

2*x_1

x_4:

x_2*x_2

+

x_1:

1.0

2 x_2:

0.5

x_2:

0.5

× ×

0.5 0.5 1.0

Memorization

function f(x_1, x_2): x_3 = 2 * x_1 x_4 = x_2^2 x_5 = x_3 + x_4 return x_5

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

40

• Backward sweep computes the derivatives from the output using the chain rule
• All partial derivatives calculated in one sweep
• Hundreds of times speedup in real-life Greek and sensitivity runs

Example: Automatic Differentiation

x_5:

x_3+x_4

x_3:

2*x_1

x_4:

x_2*x_2

+

x_1:

1.0

2 x_2:

0.5

x_2:

0.5

× ×

1 df/dx_3 = df/dx_5 * dx_5/dx_3 = 1 1 df/dx_1 = df/dx_5 * dx_5/dx_3 * dx_3/dx_1 = 2 df/dx_2 = 1 Backward Sweep

function f(x_1, x_2): x_3 = 2 * x_1 x_4 = x_2^2 x_5 = x_3 + x_4 return x_5

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

41

Example: Automatic Differentiation

• Should business users care about low-level algorithm details?

– It is unrealistic to expect business users to become proficient at algorithm design – But efficient algorithms are critical in solving today’s business problems – A system should be flexible enough to work for business users and algorithmic improvement

Data and Analytics Business Users Algorithms Computer Scientists

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

42

Example: Calculation Result (Hyper-)Cube

• Large-scale computations often result in big datasets with 3 or more dimensions (e.g. policies, steps,

scenarios, sensitivities). Outputting the entire dataset incurs excessive memory and disk overhead.

• Alternatively, the system could produce “views” of the data on demand if re-calculation is fast.

Policy Scenario Time Policy Scenario Time Policy Scenario Time Policy Scenario Time

f(x) f(x)

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

43

Conclusions

• Regulatory changes accelerate the modernization of computation systems
• Modern computation systems are fundamentally different than old computation systems

– Fast development cycle – High productivity computing through GPUs – Complete stack of software promotes enterprise-level integration – Extreme flexibility and controls without performance penalties

• Intelligent model design is challenging

– The best results are obtained through optimization of entire stack – Some optimizations can be automated but other many require expert input and judgement – Systems must be flexible enough to accommodate the latest technology and to provide users with

• ptions for tomorrow’s calculations

Prepared by PathWiseTM Solutions Group Proprietary & Confidential

44

Thank You