You are on page 1of 5

Modelling and Data

Why model are useful? How model are selected?

Usually involve making simplified assumptions. The simplified assumptions for example:

Quantity is constant over the period


We know the statistical distribution
Aspect influence the model are independent
Some aspect are insignificant
All financial instrument are always priced so as to preclude risk-free arbitrage profit

Case study: actuaries want to make an annuity product. One possible problem is the product might
provide no protection against inflation.

First step in modelling include: define the problem, find the reliable data source, building model and its
alternative.

Using a fitted model

Actuary usually uses the model to advocate the product launch, for example include:

Proposal of a pricing basis for the product, for example providing customer with real
return.
Analysis of products expected profitability
Analysis of products risk
Confirmation that the profit is sufficient given the capital required. Usually achieved by
comparing the return on capital to a hurdle rate.

Throughout the process, actuaries need to deal with:

Marketing team, to gain understanding of competitors price and feature


Operation team, to understand the product cost to be built into expense budget
Sales team, to understand the commission structure and expected target volume
Management team, to gain approval

Challenge the fitted model

Pricing committee wants to ensure the product is profitable (but pursuit of higher margin lead
to lower transaction). Usually the committee will challenge the product developer:

Interest analyzed
Whether the estimated parameter are reliable and how alternative value would affect the
modelling
Wider aspect of product design. For example: adverse selection. Super-healthy pensioner
are more concerned about long term inflation, and may find inflation linked annuity
particularly attractive (rather than choose conventional annuity). That means standard
assumption are not valid for this product.

Normative approach to modelling

1. Statistical inference, to test any hypothesis and constructing experiment to validate assumption.
2. Explanatory data analysis. Starting point of model building is collection and analyze the data. The
least formal (but important) stage is explanatory data analysis, means examining data and
looking for pattern.
3. Model calibration, means estimating the uncertain parameters in a model. The most popular
method is maximum likelihood, choosing the most likely set of parameters. Alternative approach
is Bayesian method that treat parameters as being uncertain (prior distribution capturing a
subjective view of where we expect parameter to lie).
4. Fir to evidence. With data, equations, parameters and error term we have a possible model. We
lack of evidence whether the model is correct or useful. The popular tool is stress testing. This test
is simple but will not capture likely behavior outside the data sample to which the model was
calibrated. Alternative approach is to generate several random future projections from the model.
5. Hypothesis testing. We want to see whether the model capture important aspects of historical
data. Usually involve inspection by eye rather than formal statistical test. For example: hypothesis
testing for linear regression problem:
H0 H1
Alternative
Name Null hypothesis
hypothesis
Mathematical
b=0 b<>0
formulation
Estimated b is
Estimated b is
Supportive test not significantly
significantly
result different from
different from zero
zero
Type I and II errors in hypothesis testing
Test result support Test result support H1
H0 (reject H1) (reject H0)
H0 is true Good outcome Type I error
probability 1-alpha Probability alpha
H1 is true Type II error Good outcome
probability beta Probability 1-beta
6. Parsimony. Good data fir becomes easier as more parameters are added, but adding parameters
without limit may violate the principle of parsimony. Parsimony means using only as many
parameters as necessary to fit a model.
7. Fit to theory
8. Computer development. Turning the model formulas into computer code to solve a specific
problem.
9. Using model for projection
10. Bootstrapping, is a powerful but labor-intensive test. The idea is to take a fitted model, taken to be
correct and use this, with a random number generator, to produce data sets according to that
model. This is will test many part of the modelling process.
Limitation of the normative approach

Practical difficulties
Theoretical ambiguity
Expecting the unexpected

Commercial modelling

1. The role of modelling within the actuarial control cycle. Models are designed to be used over and
over. The shelf life of a model may vary from few a weeks to many years. The model must be
updated, improved, and enhanced.
2. Cost of models and data. Include:
Cost of data collection, cleaning, processing
Cost of estimating model parameters
Cost of model testing and validation
Hardware and software cost of implementation
Cost associated with the use of model output, include communication
3. Robustness, means the model can stand up to challenge, confidence in assumptions,
mathematical formula, and software implementation. Somehow there was some limitation in a
process that make the model not that robust. But management still have to evaluate the risk. The
recipient of a modelling output tries to evaluate several things at once:
The answer to the problem posed
The quality of project management
The quality of the modelling
4. Governance and control, to manage volatility of model output. Governance is the process of
oversight in decision making.
5. Models for advocacy, many modelling project are commissioned for the purpose of advocacy,
especially in long-term actuarial project
6. Models and markets

Data

Quality of actuarial work highly depend on data (factual information used as a basis for reasoning,
discussion, and calculating) and assumptions (a fact or statement taken for granted) used. Some factors
hinder development of the best solutions are:

Lack of time
Lack of resources (computer, assistance, etc)
Lack of data/relevant data
Lack of knowledge about key factors

Reason why we need data


Present data to give accurate starting point for projecting the future
Past data to use as a guide in constructing model and setting assumption for the future
Knowledge about key factor is essential

Sources of data
1. Internal vs external, internal could be gained from internal company
2. Whole population (census) vs sampling vs survey. Whole population
3. Cross sectional vs longitudinal.
Cross sectional data arise when observation from numerous subject are collected at one
point time. Example: claim payment made in respect of automobile accident during a
fixed time period
Longitudinal data refers to observations collected over time on the same entity. Example:
stock market index
4. Obtaining high quality data. Steps in improving data quality:
Prevention: eliminating errors before they enter the database
Detection: study the collected data in a search for errors
Treatment: repair the data, dealing with errors which have been detected
5. Data checks. Item to consider in data checking:
Know where the data come from
Know why and how the data was originally captured
Understand the incentives inherent in the datas original use
Examine several randomly selected records
Have an expectation of the distribution of the data
Look for blank and duplicate
Ask for the definition of the critical data items
Develop some ways to verify the data
6. Data repair. Common method pf correcting error was return to the source and determine what
went wrong then make the correction. If the correct answer is not obvious and it is not possible to
return to the source, the alternative is imputation, usually by regressing the missing item. If
record cannot be impaired it may have to deleted or at least not used for analyses.
7. Missing or inadequate data.
8. Standard of practice and professional implications. There are few dimension of data quality:
The production of data should be impartial and objective
Relevance: data should be adequate in scope of coverage
Timeliness: delay between the events measured and when the data become available
should be not too great.
Accuracy: sampling error is the most critical components of actuarial studies
Coherence: there should be internal consistency in the data as well as consistency with
regard to previous studies and those by others
Interpretability: data should lead to result which can interpreted by and be meaningful to
users
Accessibly: the data and reports must be available to those who will use them
9. Challenge presented by limited data. For example: liability insurance claim such as medical
malpractice are fairly infrequent. This is particularly problematic when setting assumptions such
as selecting a probability model. Having ten or twenty observation can make this very difficult.

Assumptions

Control cycle process for setting assumptions:


The assumption- setting control cycle
Steps for assumption setting:
Identify the assumption: list of assumption required
Quantify the assumptions: assigning numerical value
Monitor the assumption: as experience is obtained, it can be used to make appropriate
numerical changes to existing assumptions

Identification of assumptions.
Sometimes it is not so obvious about what assumption should be used. Categorization can help to
identify assumptions. One of the categorization is by source of data used to set the assumption.
Source can include historical experience from:
The same company and product
The same company but different (but similar) product
Similar product sold by other companies
Source unrelated to the company or product

Quantifying assumptions
Actuarial work often uses what is called the best estimate, at least for a starting point. It
usually considered to be the expected value.
Then from that best estimate, prudent estimate can be reached by making explicit
additional margin for uncertainty.
The materiality of an assumption refers to how much impact a change in the
assumption will have on the eventual result. Sensitivity can be used to evaluating
materiality.

You might also like