non-traditional actuarial work

22
Non-Traditional Actuarial Work SCCAC - 2016/11/30 Frank Chang CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Upload: others

Post on 05-Dec-2021

2 views

Category:

Documents


0 download

TRANSCRIPT

Non-Traditional Actuarial WorkSCCAC - 2016/11/30Frank Chang

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Traditional Actuaries (P&C)

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

● Pricing / Product Actuaries: Estimate insurance premiums, may engage in some research, tool-building, or product development

● Reserving / Corporate / Financial Actuaries: Estimate loss reserves, reserve uncertainty, may be involved in reinsurance, capital allocation, statutory accounting

● Research / Modeling Actuaries:Build cat models, class plans, claims analytics, credit models, gain competitive advantage / efficiency

Employed by Insurance Companies, Consultancies, Regulators

Non-Traditional Actuaries (P&C)

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

● Pricing Actuaries: Estimate cost of risk, optimal retention (ERM), impact of operational decisions, insurance purchasing decisions

● Reserving Actuaries: Estimate self-insured or captive loss accruals and expense, may model VaR, tVaR

● Research Actuaries:Construct predictive models to mitigate or allocate loss, increase operational efficiency

Employed by Other Companies

Non-Traditional Actuaries (P&C)

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

● Product Management: Estimate pricing / structure / design for company’s insurance-related product offering

● Risk Management: Decide on insurance purchasing, risk allocation, distribution within risk-bearing entities, manage risk profile

● Data Science / Statistician:Conduct experiments, provide data-based results to guide decision-making, may be embedded or centralized

Employed by Other Companies

I. Bootstrapping:Overview

Simple Generalized Linear Model (ODP)

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

1. Set i = accident year, j = development age2. Inflation factor, going down accident years: xi3. Development factor, going across maturities: yj4. k, a constant

A few math slides to get us to the bootstrap!

Set Up Model

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

1. Set mij = incremental loss for year i at age j

2. Find xi and yj to best fit the data: mij = kxiyj

3. Take the log of mij = kxiyj to make it easier to fit

ln(mij) = c + αi + βj

where αi = ln(xi) and βj = ln(yj)

A few math slides to get us to the bootstrap!

Analyze Results

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Use generalized Pearson residuals to judge fit:- uniquely defined for negative incremental loss- calculated consistent with scale parameter

where

95% of residuals should lie inside (-2, 2)

A few math slides to get us to the bootstrap!

Reduce, Reuse, Recycle

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Use residuals to bootstrap triangle

Bootstrap Results

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Output from “Chain Ladder” package

Gives range for estimates for

Pricing: Ultimate cost

Reserving: “IBNR”

II. Application ofBootstrapping

Operational Planning : Handsets

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Unit Sales Forecast Process

Demand

1. Weekly Forecast

Demand

Demand

By unitBy region

2. Incorporate Customer Input

Carriers

3. Global Shifting

MOTO

4. Supply-Demand Management

Adjustment AdjustmentDemantra

or equivalent

Operational Planning : Bootstrap

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Organize Data

Operational Planning : Bootstrap

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Organize Data - Triangular Form

● What patterns do we see?

● What questions can we ask?

Operational Planning : Bootstrap

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Calculate Model Parameters, Residuals

● Construct Forecast / Actual Ratios● Calculate residuals by vintage, segment

Operational Planning : Bootstrap

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Results

● 2013 YE financial forecast of 3,050 (75th %ile)● 2013q1 actual: 2,850 units (close to mean)

III. Overview of Big Data

Tools Used in Actuarial Research

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Overview

Non-Distributed Computing:● R

● Python

Distributed Computing:● Hive

● Spark (Scala / Java / Python / R)

Tools Used in Actuarial Research

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

R / R Studio / R Studio Server

● Software with stats-focused libraries

● Handles more data than Excel (up to 5 GB*)

● Great visualization libraries (e.g. Shiny)

● Not really a programming language

* Revolution R (now owned by Microsoft) can handle more

Tools Used in Actuarial Research

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Python / iPython (juPyter) Notebook

● Well-supported programming language (2.x)

● Fast, great for ETL (pandas)

● Handles bigger datasets (up to 40 GB)

● Not stats or visualization focused

● Easy-to-learn, easy to share work (notebooks)

Tools Used in Actuarial Research

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Apache Hive

● SQL-like interface to Hadoop

● Distributed Data (limited only by hardware) ● ETL Tool to get data into manageable form

● Bad max speed (set up mappers and reducers)

● Min speed is determined by worse executor

Tools Used in Actuarial Research

CONFIDENTIAL - DO NOT COPY, CITE, OR DISTRIBUTE WITHOUT PERMISSION OF THE AUTHOR

Apache Spark

● Tool with 4-language interface

● Distributed Data (limited only by hardware) ● Can write libraries to be incorporated as jars

● Better max speed (all work done in memory)

● Can handle ETL as well as statistical modeling