Escolar Documentos
Profissional Documentos
Cultura Documentos
Optimising Compliance
The role of analytic techniques
__________________
Decision Tree audit-csv.txt $ Adjusted
Marital
2 Education
0
755 cases
93.8% Occupation 7
1
219 cases
12 Hours < - > 38 74.4%
0
322 cases
73.6% 26 Age < - > 33.5
0
21 cases
71.4% 54 55
0 1
22 cases 61 cases
63.6% 73.8%
__________________
Abstract
The use of quantitative analytical techniques has an important and growing role in the optimisation of revenue
authority client compliance. This paper covers some of the issues associated with the application of analytic
techniques in a revenue authority business setting.
The views expressed in this paper are those of the author and do not necessarily reflect the considered views of
my colleagues at the Australian Taxation Office on any matter.
Stuart Hamilton
Assistant Commissioner
Corporate Intelligence & Risk
Australian Taxation Office
2 Constitution Avenue
Canberra ACT 2600
Overview .............................................................................................................................3
Background
The ATO intent and business model ....................................................................................3
Analytics and the personalisation of interactions .................................................................5
Scene setting
Non-compliance ...................................................................................................................6
Strategic Risk.......................................................................................................................7
A consistent measuring framework for Revenue Risk..........................................................9
Optimisation
Treatments available (Enhance treatments) ......................................................................17
Assignment of clients to those treatments (Enhance candidate selection process) ...........20
Case mix (Enhance case mix) ...........................................................................................26
Other matters
Prioritising analytic work.....................................................................................................29
Measuring strikes ...............................................................................................................30
The effect of varying strike rates ........................................................................................30
Traditional versus analytic driven case selection – an example .........................................33
Optimising capability ..........................................................................................................34
Conclusions .....................................................................................................................34
Annexes ............................................................................................................................35
Scoring using ∆Tax ............................................................................................................35
Some additional Rattle analysis output ..............................................................................36
Spreadsheets used in this paper........................................................................................37
Summary of some analytic methodologies that might be used in optimisation ..................38
Overview
The use of quantitative analytical techniques has an important and growing role in the optimisation of revenue
authority client compliance. This paper covers some of the issues associated with the use of analytic techniques in
a revenue authority business setting.
The ATO Business Model is premised on clients self assessing their tax obligations, with the ATO providing
education and assistance, making it easier to comply, and verifying that compliance is occurring using appropriate
risk management approaches.
Addressing non-compliance in an appropriate manner requires use of the ATO the Taxpayers’ Charter principles
regarding taxpayer rights and obligations; the Compliance Model view of treatment selection and escalation, and
the Brand Navigator on the appropriate ATO persona to present.
2
These views are brought together in the ATO Community Relationship Model :
1
See http://www.ato.gov.au/content/downloads/ARL_77317_Strategic_Statement_booklet.pdf
2
See http://www.ato.gov.au/docs/CommunityRelationshipModel.doc
The “Easier, Cheaper and More Personalised” change program sets out eight key principles to guide the design
and development of ATO products and services. Two of the guiding principles link to the services that are enabled
or facilitated by analytic methods.
3
See http://www.ato.gov.au/content/downloads/Making_it_easier_to_comply_2005_06.pdf
This is the essence of analytics, being able to examine large data holdings to identify those attributes that correlate
to particular client circumstances, behaviours and risk profiles. Those attributes can then be used to enable the
ATO to better tailor its client interactions. That is to make the interaction more relevant to the client – more
personalised, more likely to succeed.
This examination of client attributes to identify those that correlate to particular client circumstances, behaviours
and risk profiles is an area of analytics known as ‘data mining’ or knowledge discovery and it is this approach that
heralds an improved way of optimising client compliance. Before delving into these analytic approaches some
scene setting is needed to establish a consistent base of terminology and understanding for this paper.
Economic theory (and common sense!) tells us that there is an optimum point in a revenue system at which the net
revenue of the system would be maximised (Os) for given tax rates and rules. Beyond that point the marginal cost
exceeds marginal revenue – it costs more than a dollar to get a dollar in.
In most economies the revenue authority would operate significantly below this point (O1) reflecting lower
budgetary needs of the Governments, administration resource constraints and accepted community attitudes
regarding the nature and level of intrusion of the revenue authority into the economy.
$ Theoretical “Full Compliance” bandwidth
Total
Revenue
Total
Cost
Resource Theoretical
Constrained System
Optimum Optimum Net
Revenue
O O Cost
In these circumstances1 “optimisation” takes on as different practical meaning.
The challenge for the revenue authority is to optimise long term net revenue outcomes within its resource and other
operating constraints – ie to maximise long term voluntary compliance given relatively fixed resources.
4
Compliance action in accordance with the ATO Compliance Model and the Taxpayers’ Charter. “The compliance model
directs that we better understand why people are not complying and that we develop appropriate and proportionate responses. An underlying
objective is to develop responses that maximise the proportion of the community who are both able to, and choose to, comply.” Depending on
the reasons for non-compliance, responses can be aimed at: enabling compliance through education, assistance or making it simpler to comply,
or enforcing compliance through via administrative and prosecution action.
In order to optimise long term voluntary compliance it is suggested that number of aspects need to occur:
o The number of clients requiring relatively expensive remedial compliance action should be minimised (ie
clients should know how to comply [education], be able to comply [the interaction of client and the revenue
authorities systems needs to 'fit'] and be ready to comply [attitudinal]).
o The right range of remedial compliance treatments should be available. Any remedial compliance action
should maximise the long term compliance gain within the client base (ie the treatment should be most
appropriate to engender voluntary compliance overall – a leverage compliance model viewpoint).
o The right clients should be selected for the appropriate remedial compliance action (ie the strike rate for a
particular remedial compliance treatment should be maximised).
o The right mix of discretionary compliance work needs to be achieved overall so that long term revenues are
optimised within the resource and capability constraints faced by the organisation.
Revenue authorities generally have a range of administrative treatments: education, assistance, review and
enforcement products that they can use to address non-compliance, in addition to generally longer term system
and legislative/policy changes.
At present optimisation in tax administration is generally based on the expert views of experienced senior officers.
Equally valid, but different views might also be held and hence the outcome is one of subjective optimisation.
Colloquially it might be put as ‘informed gut feel’ of the right balance - a judgment call.
It is suggested in this paper that these judgment based optimisation approaches may be better informed and often
enhanced by the use of objective business decision support approaches that assist in the determination of the
‘right’ products, the ‘right mix’ of products and the ‘right’ clients to apply those treatments to.
However these objective approaches do require reasonably robust knowledge of the costs and benefits of various
courses of action and this can be a significant problem. A greater level of consistency of approach and
5
understanding is also needed so that the information used in business decision support is valid and verifiable .
While there are many aspects to compliance with revenue laws and regulations (see the ATO Compliance
6
Program) client obligations can generally be thought of in the following broad manner :
o Registering in the system (either with the revenue authority or with some other body)
Most revenue systems also require a client to maintain records of appropriate information for some set period. Ie
o Keeping records that allow verification of the information used to satisfy the above obligations.
5
If such information is not available then simulation modelling can still assist in determining key directions and sensitivity aspects.
6
See page 7 of the OECD 2004 document: Compliance Risk Management @ http://www.oecd.org/dataoecd/44/19/33818656.pdf
The IRS Tax Compliance Measurement Program (now known as the National Research Program -NRP) is primarily
based on a large scale periodic stratified random audit approach and indicates that:
With an overall non-compliance rate in the USA of ~15% of the theoretical tax believed due.
Simplistically, if the voluntary compliance cost and response was equal between these types of non-compliance
then the compliance resource allocation should broadly match the above splits. This simplistic view does not take
into account that resource intensity and compliance responsiveness will in practice differ significantly depending on
the nature, causes and extent of the non-compliance and the available treatments for it. (More on this later.) Such
a macro or high level view is of limited practical use in understanding and addressing non-compliance.
7
Source: http://www.irs.gov/pub/irs-news/tax_gap_figures.pdf
Moreover, while superficially compliance may seem a strict matter of fact – yes or no - in practice it is often far
8
more blurred or ‘grey’, a question of interpretation and judgement, than may appear to a layman . Uncertainties in
definitions, interpretation and measurement compound to render views on what full compliance is as a relatively
broad bandwidth with wide confidence levels. Movements in total compliance are very difficult to ascertain with a
high degree of confidence and are necessarily dated with this periodic sampling approach.
7
Note - the figures for underreporting are more subjective as they are inflated (by a factor between 2 and 3) to take into
account income not detected by audit processes. See for example the US Tax Inspector General for Tax Administrations April 2006 report @
http://www.ustreas.gov/tigta/auditreports/2006reports/200650077fr.pdf#search=%22US%20IRS%20Tax%20Gap%20Estimate%20Inspector%2
0General%22
8
See the OECD 1999 document: Compliance Measurement at page 4 @ http://www.oecd.org/dataoecd/36/1/1908448.pdf
“To sum up, the general position on measuring the tax gap is that it is difficult if not downright impossible and even
if it were possible to get a reliable total figure it would not tell us much of practical value in the struggle against non-
compliance.”
That does not mean that a random audit program such as that use by the US IRS NRP does not have its place in
the optimising compliance tool kit. It just needs to be used appropriately – when knowledge of the risks of a client
group is relatively low so that the low strike rate, and hence high rate of intrusion into compliant clients affairs, of a
significant random case selection approach can be justified overall in terms.
Broadly speaking when optimising compliance I would suggest that any large scale random audit program be
restricted to those situations where the true informational value and deterrence effects of the random process
outweigh the significant social and opportunity costs it imposes over targeted detection based approaches. Ie The
few areas where the organisation is flying blind. Eg Where new legislation has been introduced and the
discriminate features of non-compliance are largely unknown and can’t be reasonably estimated.
o Risk based approaches are appropriate where we can target clients by way of known risks. (Note that we
can still discover new risks to the extent they occur, and are detected, in the clients reviewed.)
Low High
Knowledge of risk in client population
If you have a robust and effective measure of client risks and how to detect them then discovery processes
can be kept relatively smaller than the situation where you have little or no knowledge of client risks.
Discovery processes often have a low strike rate – the value is generally in the insight into better detection
approaches.
Note: The volume of cases needed to discover relative differences in risk (and so tune selection
approaches) is significantly lower than the volume of cases needed for a statistically accurate estimate of
the overall tax gap.
Most leading OECD revenue authorities have a reasonably robust knowledge of their client base and the broad
strategic risks do not generally change much year to year. Overall most of their clients pay most of their tax most
of the time. The ATO's compliance program http://www.ato.gov.au/content/downloads/ARL_77362_n7769-8-
2006_w.pdf details the ATO view of the compliance risks and issues facing the organisation, the treatments being
applied and the results.
So if we don't use a tax gap analysis to inform the relativities of strategic and operational intelligence (the what and
who to treat) is there a consistent, objective framework that can be applied across products and obligations?
Without an objective mechanism our legal 'choice of remedy' may be more open to criticism and question over
concerns of bias, subjectivity and inconsistency. (Why did we select case X over case Y or risk A over risk B?)
Broadly speaking the following conjectures might be stated regarding the requirements of a revenue risk
measurement approach:
o For the client obligations of registration, on time lodgment, accurate reporting and correct
accounting/payment, client revenue risk should be consistently viewed where this is practical.
o Client revenue risk ranking should be done as objectively as possible, taking into account aspects such as
the revenue at risk, the severity of non-compliance and our confidence level in the revenue at risk.
o Client revenue risk ranking should be flexible enough to cater for aspects such as losses, recidivism,
schemes and associated client linkages.
o Client revenue risk ranking should be highly scalable and stable – from lists of a few clients to potentially
scoring millions of clients.
For the client obligations of registration, lodgment, accurate reporting and correct accounting/payment, I suggest
that revenue risk can generally be distilled into the following four features for the purposes of comparable revenue
risk ranking across and within obligations and products:
o ∆Tax Delta tax - The change in primary tax associated with the non-compliance.
[ie Identifies those who may have the most tax wrong. An absolute amount. “Client A may have underpaid
$5,500 in tax in year y.”]
o ∆Tax/(∆Tax + Tax) Severity - The relative severity of the non-compliance as a percentage of tax paid.
[ie Identifies those who may have most of their tax wrong. A relative value. “Client A may have underpaid
15% of their tax in year y.”]
o Cf(∆Tax) Confidence - The confidence interval associated with our estimate of ∆Tax. [ie
Identifies how confident we are of the estimate in ∆Tax. “We are 90% confident that Client A underpaid
$5,500 in tax in year y +/- $550”]
Let us look at how this revenue risk concept might work for the key client obligations previously identified:
o Registration
For registration non-compliance ∆Tax is the estimate of the amount of net revenue that is predicted to be
derived by achieving registration (and subsequent lodgement) for those clients not registered in the
system. (Detection is generally via matching processes, comparing a list of names against registered
clients to detect those operating outside the system.)
o Lodgment
For non Lodgers: ∆Tax is the estimate of the amount of net revenue (ie taking into account withholding and
instalments: PAYGW & PAYGI etc) that is predicted to be derived by achieving lodgment from a treatment
for those clients who otherwise would not lodge.
Late Lodgers: For those clients who would lodge late ∆Tax is really the time value of money (PV) brought
forward by achieving earlier lodgment following a treatment. Most lodgment clients by number would be
late lodgers rather than non lodgers. The distinction between the two classes will overlap and an informed
decision is needed for the transition from one class to the other.. Eg After 3 months a late lodger
'transitions' into a non lodger.
This is of concern as in mature tax systems most non-compliance is likely to be associated with
understatements of taxable income or turnover (or the over claiming of deductions or credits).
o Correct accounting/Payment
Non Payers: ∆Tax is the estimate of the amount of net revenue (ie taking into account PAYGW & PAYGI
etc) that is predicted to be derived by achieving payment from a treatment for those clients who otherwise
would not pay.
Late payers: For those clients who would pay late ∆Tax is the time value of money (PV) brought forward by
achieving earlier payment following a treatment. Most debt clients by number would be late payers rather
than non payers. As with the late/non lodgers the two classes will overlap and an informed decision point
is needed for the transition from one class to the other.
Having grouped our non-compliance using a common revenue measuring stick of ∆Tax, what are some of the
salient aspects we would expect to see in the resultant distribution?
Expected ∆Tax distribution where ∆Tax is the estimated primary tax involved in the non-compliance:
o Modal ~ 0. ie most clients comply with their tax obligations (in accordance with the compliance model
view.)
o Smaller negative tail. ie clients, and their advisors, have more of an incentive to detect, or not to make,
errors in this direction.
o Longer positive tail following a truncated pareto or power distribution past $Y. ie most errors or omissions
are relatively small while a few are very large. The fall-off is basically in accordance with the well known
pareto distribution (a form of inverse power distribution).
o Our ability to predict ∆Tax from data is limited and has a confidence interval associated with it. Our
confidence will be greatest where we have significant experience and falls off as we move away from this.
o Risk scores based on deviations from average ∆Tax would be relatively stable for large populations. Ie the
Average overall tax underpaid would not vary significantly from year to year.
xx
x
x
x x
Confidence distribution
x in ∆Tax Estimate
x x
x
x x x
x x
x x x
x
x x x x x x
x x x x x
x x
x x x x x x
x x x
x x
x x x x x x
x
∆Tax →
$Y
This gives a indication of the relative ‘severity’ of the non-compliance taken as a proportion of the tax paid:
0 implies the amount of the error, avoidance or evasion is relatively insignificant compared to the tax the
client has paid.
1 implies that the tax error, avoidance or evasion affected all or almost all of the tax otherwise payable.
o Smaller though longer negative tail clients, and their advisors, have an incentive not to make errors in
this direction, but an overpayment may exceed the total tax that would have been due.
n
x
x
x Confidence distribution in
x
x x ∆Tax/(∆Tax+Tax) Estimate
x
x x
x
x x
x x x
x x
x x
x x x x
x x x x x x
x x x x x
x x
x x x x
x
x x
x x
x x x
x x
0
∆Tax/(∆Tax+Tax) 1
These factors – the absolute amount, the relative amount, our confidence in the estimate and our view of
collectability are important in objectively prioritising our relative concept of risk so that we can optimise the selection
of treatments and clients.
A balance must be struck by the organisation so that case selection prioritisation is defendable and repeatable (so
that our ranking of clients can be evaluated and improved.)
This approach of estimating client errors also raises the aspect of identifying and dealing appropriately with
overpayment or credit situations – both in an absolute and relative sense.
n
xx
x
x x Confidence
x distribution in
x ∆Tax Estimate
x x x
x x x
x x
x x x x x
x x x x x
xx x xx x
x x x xx x x x x
x x x
x x x x x x
x
$Y ∆Tax →
While a ∆Tax/(∆Tax + Tax) estimate gives us a view as to who may have evaded or avoided most of their tax in
relative terms – a critical factor for a revenue collection agency looking at serious non-compliance and aggressive
tax planning.
n
xx
x
x x Confidence distribution in
x x
x x ∆Tax/(∆Tax+Tax) Estimate
x x x
x x x
x x x
x x x x x x x xx x
x x x x
x x x x x x
x x x x x x
x x x x x
x
0
∆Tax/(∆Tax+Tax) 1
o Ranking
Ranking scores could be produced for both revenue, severity, confidence and collectability:
A revenue score is produced by dividing the client ∆Tax by the average ∆Tax [ie ∆Tax/Ave∆Tax].
A severity score is produced by dividing the client severity [∆Tax/(∆Tax + Tax)] by the average severity. (If
required high, medium and low risk classes could be constructed in regard to the distribution – high risk say
the highest 10%, medium risk the next 20% and low risk the next 70%.)
This approach to revenue risk scoring can provide an objective relative ranking of tax risks and cater for issues
such as recidivism, losses and the promotion or association of non-compliance by others.
o Losses
For multi year scoring losses can be accommodated via the PV of the tax adjustment claw back over a
standard time horizon (say three years).
o Agents
Agents could be looked in regard to both the ∑ of the ∆Tax of their clients and also from a relative severity
approach. Relatively high scoring agents could be assigned for an appropriate treatment.
o Whole of Client
The values of estimated ∆Tax can be summed for a client provided they derive from mutually exclusive
risks or they are counted on a first point of error basis. (ie Work related expenditure ∆Tax and rental
property ∆Tax and omitted income ∆Tax can be summed whereas ∆Tax for WRE and over-claimed
expenses ∆Tax cannot (unless over-claimed expenses is constructed to exclude WRE claims)).
A whole of client / whole of product risk profile can then be constructed that places the clients and product revenue
risk in relative relationship to other clients / products.
In the ATO these broad client obligation risks are now expanded upon in the case management system down to
return form/schedule label items.
The rationale for this is that incorrect reporting ultimately relates to aspects not correctly reported at a label item.
Our tax return database consists of these label item values by form by client by reporting period and is generally a
critical source of discriminate information used in case selection processes.
Risk Score Risk Score Risk Score Risk Score Risk Score
Incom e Tax
GST Risk Score Risk Score Risk Score Risk Score Risk Score
Excise Risk Score Risk Score Risk Score Risk Score Risk Score
Super Risk Score Risk Score Risk Score Risk Score Risk Score
(other FBT etc) Risk Score Risk Score Risk Score Risk Score Risk Score
All Products
(weighted scores) Risk Score Risk Score Risk Score Risk Score Risk Score
*NOTE: Client Scores can be further aggregated to support Industry, Occupation and Product Risk Scores for the w hole client population. This ability is critical to realising a more robust risk assessment
process and linking corporate risk rating w ith case-based risk rating.
It should be noted that the summation of ∆Tax cannot be used to construct a tax gap view like that of the IRS:
Why ∑∆Tax cannot be used for Tax Gap estimates…
The ∆Tax predictions are derived from the analysis of successful cases (strikes) and are not drawn from a
representative sample of the population. At low ∆Tax, predictions the confidence interval would be
relatively large compared to ∆Tax. Ie The uncertainty in the tax revenue at risk would exceed the estimate
of the revenue involved – hence it cannot be used to project overall compliance levels accurately.
Diagrammatically: n
xx
x
x x
x
Confidence distribution
x in ∆Tax Estimate
x x x
x x x
x x
x x x
x
x x x x x x
x x x x x
x x
x x x x x x
x x x x
x
x x x x x x
x
0 ∆Tax →
However ∆Tax predictions could be used to assist budget revenue at risk as they are based on past claw
back.
It is suggested that a better view of revenue risk is the risk to budgeted revenue flows from non-compliance rather
than total tax gap concepts. The sum (∑) of ∆Tax from successful cases (strikes) is the direct claw back from
compliance activities included in the consolidated revenue received by Government and built into budget estimates.
Estimates of ∆Tax and changes in ∆Tax could assist in the annual risk analysis process undertaken each year that
is used to guide mitigation activities.
Revenue Consequence
Distinguish betw een compliance behaviour changes and economic changes
beyond our influence of control
Revenue
Actual Forecast
∑∆Tax
More revenue
More revenue Predicted
Budgeted v claw back
∑∆Tax of tax
Revenue Actual
claw back
of tax Less revenue
Less revenue
Predicted
claw back
of tax
Before turning to the use of analytic techniques for optimisation we should note that it is crucial that the qualitative
information from our intelligence gathering activities be combined appropriately with the quantitative intelligence
produced via the use of analytics. Bring both views together produces a better result than using just one.
Intelligence views:
“Data! Data! Data!" he cried impatiently. "I can't make bricks without clay." Sherlock Holmes in ‘The Adventure of
the Copper Beeches’ (1892)
Having set the scene and established a view of what non-compliance is and how we might consistently measure
revenue risk, we now look at how we determine the right treatment for the right client.
Identify
strategic risks
Select
risk to be
treated
Define
candidate Enhance
population treatments
for risk
Enhance
candidate
selection
process
We will then examine approaches for case mix optimisation – how does the mix of case types and numbers impact
upon the revenue and can that be optimised.
Systems where the only answer is a prosecution will tend to view the solution set to a compliance issue as
prosecuting the right clients – even though a prosecution might not be the ‘right’ treatment to engender long term
compliant behaviour. (Indeed a side effect can be that an enforcement culture permeates the organisation rather
than a client service ethic that realises clients get it wrong for a wide variety of reasons.)
Research on regulatory compliance models indicates that a range of treatments should be available to engender
long term voluntary compliance. An escalatory model is suggested to create an incentive for the client to move
towards a more engaged compliant behaviour set. This should take into account the facts and circumstances of
the clients’ situation so as to treat the client in the most appropriate way:
o For example recidivist clients (those who repeatedly offend after treatment) would generally warrant a
different treatment than a client detected making an error for the first time.
o Similarly those who promote non-compliance by others generally warrant different treatment to those who
don’t.
o Those in special positions of trust and influence in the tax system (eg key intermediaries, revenue authority
staff, lawyers and accountants) generally warrant different treatment to those who aren’t in such positions.
o Those involved in avoidance schemes aren’t all the same. Clients with relatively low knowledge of the tax
system who enter schemes on the advice of their trusted advisor should not be treated the same as those
who would be reasonably expected to have good knowledge (such as the advisor).
The causal factors involved in the non-compliance also factor into the appropriateness of the choice of remedy.
o a difference of views – a reasonably arguable position that differs from the ATO view,
o honest mistake,
o ignorance,
o carelessness,
o negligence or
o deliberate intent.
It is important that our 'choice of remedy' be appropriate and defendable – and that the mechanism to get to the
decision on the remedy be evidence based and repeatable.
These aspects can be brought together into an overall ‘model’ or framework to view non compliant behaviours
based on the clients level of engagement with the regulatory system.
As the NZ IRD version of the compliance model indicates (See diagram on next page):
o Some treatments will apply to many clients at once and be rather general in nature – such as education
materials available to the public.
o Other treatments will be client group specific, advice aimed at a particular industry or occupation or
segment.
The New Zealand IRD model is a useful adaptation on the revenue authority compliance model presented in the
1998 ATO Cash Economy Task Force Report:
o For those trying and succeeding to do the right thing – the majority of clients – compliance is made as
simple as possible. Information requirements are reduced and interactions are made as cheap and easy
as is practical. (See for example the ATO Easier, Cheaper, More Personalised Change Program @
http://www.ato.gov.au/content/downloads/Making_it_easier_to_comply_2005_06.pdf )
o For those trying, but not succeeding, in doing the right thing, education and advice is provided. This can be
general, or aimed at specific client segment – an industry or occupation group or some other discernable
client grouping. (See http://atogovau/corporate/content.asp?doc=/content/42628.htm on marketing and
taxation. Here descriptive analytics to identify market segments can assist.)
o Some clients may request assistance and advice and others may be targeted for a review whose outcome
is advice eg a record keeping review. Here predictive analytics can assist.
These interventions are generally relatively low cost mechanisms for enhancing voluntary compliance for clients
who are trying to do the right thing.
o A smaller number of clients will usually exist who for a variety of reasons appear to have carelessly,
negligently or deliberately not complied. For these clients a common treatment is to audit the client to
determine the amount of the non-compliance and the reasons and, if considered appropriate, penalise the
client for not complying. The audit may be targeted at a specific issue or may be a more wide ranging
examination of the whole of the client’s tax affairs.
o For those few clients that have relatively serious/aggressive non-compliance and other aggravating factors,
the treatment may be to investigate with a view to prosecution. Due to the legal evidence gathering nature
of these cases they tend to relatively resource intensive and costly.
In order to objectively optimise compliance treatments for clients we need to capture data that reflects relevant
client circumstances, the nature of the treatments used and the clients’ response to the intervention over time.
Rather than sequentially changing our treatments over time we can evolve our optimum (champion) treatment via
the use of controlled champion / challenger treatment groups where analytically similar clients are assigned to
different (though still appropriate!) treatments in order to evaluate which treatment works best at engendering long
term voluntary compliance for the relevant client segment.
By creating control groups of clients assigned to different treatment pools, champion / challenger strategies can be
evaluated and used to determine the best treatment for a client given their facts and circumstances as revealed in
the data.
To work best this needs to be a deliberate strategy with controlled data capture rather than an after the event
thought when much of the required data is no longer available for analysis.
Standard parametric (or non parametric if the control groups are small) statistical tests can be used to identify the
optimum treatment from the options tried.
Champion
Today
Potential
actions
Challenger 1
Break even
TIME
Ideally supervised learning approaches (predictive analytics) can be used to identify and assign clients into the
most appropriate segment given their particular facts and circumstances as revealed in the data holdings by an
analysis of discriminate features in the data.
Descriptive analytic techniques (eg clustering) can help you to understand the circumstances of groups of clients
that have difficulty in complying while supervised learning approaches 'mine' data on past interactions to discover
knowledge (associations – patterns and trends) that allow us to better predict the response of a client to an
interaction.
Dataminer’s use statistical and machine learning algorithms squeeze the maximum informational value out of the
dataset in question. Often the datasets have to be transformed and cleansed to get the most out of them. Some
statistical techniques for example work best with normally distributed, homoscedastic data.
Much of the 'art' of datamining stems from the knowledge and experience of the dataminer in recognising what
technique is most appropriate to improve the signal to noise ratio (ie the identification of patterns in the data) in a
given circumstance.
With supervised learning approaches the algorithms try to identify discriminate features in the data that relate to a
target variable. (A discriminate feature is one that is positively or negatively correlated to the target variable in a
statistically significant manner.)
For example clients that are considered at high risk of not responding to a letter or telephone call, because of past
recidivism in relation to a matter as revealed by data on past interactions, may be assigned directly to a team for a
review rather than to the lower cost work stream of an automated letter.
Some data mining algorithms can be conceptually quite simple, such as rule induction and decision trees, through
to the more difficult to explain neural network (essentially a black box learned weighting approach), support vector
(a machine learning partitioning approach) and random forest approaches (building a large number of decision
trees from random samples of the data that then vote on the classification outcome).
Descriptive and predictive analytics are a relatively recent field of business intelligence, enabled by the
convergence of computing power, machine learning approaches and statistical understanding.
Within the ATO we have a core group of ~16 data miners (most with PhD's in quantitative sciences), who use a
variety of software tools such as SAS Enterprise Miner and SAS JMP, NCR Teradata Warehouse Miner and more
recently open source tools like the excellent 'Rattle' to assist them in their work.
I'll provide some examples from the use of Rattle, written by one of our senior data miners to harness a variety of
open source statistical packages available in R, to give a feel for some of the advanced analytic techniques now
available in easy to use software. Rattle is available from http://rattle.togaware.com/
Exploring data relationships with Rattle – to identify possible features for selection modelling
Proportion <= x
Proportion <= x
0.8
0.8
0.8
0.4
0.4
0 0 0
1 1 1
0.0
0.0
0.0
Rattle 2006-10-04 12:49:20 Stuart Rattle 2006-10-04 12:49:20 Stuart Rattle 2006-10-04 12:49:20 Stuart
Private
Proportion <= x
Proportion <= x
0.8
0.8
Consultant
Adjusted Adjusted PSLocal
NA's Adjusted
All All
0.4
0.4
PSState All
0 0 SelfEmp 0
1 1 PSFederal
1
0.0
0.0
Unemployed
Frequency
Rattle 2006-10-04 12:49:20 Stuart Rattle 2006-10-04 12:49:20 Stuart Rattle 2006-10-04 12:49:21 Stuart
Some more examples of the analysis methodologies supported in Rattle are contained in an Annex to this paper.
Descriptive analytics.
Segmentation by decision tree against target variable (adjustment).
Marital
2 Education
0
755 cases
93.8% Occupation 7
1
219 cases
12 Hours < - > 38 74.4%
0
322 cases
73.6% 26 Age < - > 33.5
0
21 cases
71.4% 54 55
0 1
22 cases 61 cases
63.6% 73.8%
Page: 23
1121
1488
1220
1414
1007
563
49
1025
205
495
1995
183
733
1226
1537
224
130
1181
573
604
1055
1015
482
457
818
189
1194
1467
728
1255
1810
1586
1892
1711
1783
575
1095
1327
627
1715
1962
988
1726
344
822
1428
1607
784
459
1454
1170
1219
1408
777
230
865
548
232
1489
1091
1361
792
556
991
787
234
418
570
1654
1260
622
1054
1977
386
1147
1517
642
1229
1037
105
292
1852
1837
695
1538
1999
115
1472
1970
1581
1119
602
751
1196
1653
860
164
1708
462
515
1492
219
13
1817
1480
746
1941
1770
569
1744
1018
1003
74
1451
765
879
1269
560
815
1782
1353
619
797
649
69
1762
2000
1364
178
1659
204
917
1539
1655
826
1358
1650
1177
265
895
27
1210
1598
1859
217
1312
1978
193
1017
1683
1225
198
729
53
1777
1024
493
1975
1920
1354
992
1267
315
567
330
177
615
1039
428
400
1253
766
341
1519
1179
679
64
1916
1053
1491
220
149
401
911
162
184
1522
1689
611
1410
1045
1384
45
1143
1976
632
1125
1295
1592
1415
1388
626
1571
1851
350
1577
1789
138
179
1189
361
109
1356
650
696
1.5
1634
1839
126
1479
1904
431
1701
1118
1466
1074
1772
1575
176
1827
660
120
1732
390
297
1287
1847
1368
1611
403
883
811
1795
1759
693
451
371
758
1030
1335
216
1719
262
107
967
1693
1685
252
1105
445
522
1826
377
75
1440
1155
119
1887
1779
239
458
1086
1555
708
1032
1066
237
1416
1969
95
1674
320
763
1447
1982
250
537
379
118
275
638
949
31
1960
1094
1516
543
267
163
1953
1117
274
245
1430
1946
781
1661
853
89
1165
546
606
1425
1049
1972
225
631
100
946
1908
831
668
1573
1340
1075
1820
724
52
1111
1766
340
1824
1664
809
678
1063
1919
639
943
789
748
1968
623
266
1934
1008
159
1469
1619
738
578
1251
1130
1860
1679
1128
754
1434
1831
30
1706
502
165
1379
1735
211
783
1280
1518
576
1507
1604
888
1681
1159
354
1288
1709
288
1534
1411
1665
1242
1035
1216
1383
1011
114
210
1402
169
388
1192
1464
531
1332
1853
1380
610
1040
448
24
1811
1266
1221
1657
527
717
1833
1442
735
898
251
923
760
1882
1041
1231
1254
442
368
1944
1930
785
1138
1163
1197
1974
1286
1530
385
1044
1099
374
1387
1696
4
1529
812
66
1895
549
601
1281
1067
1319
595
1594
721
58
1774
1935
753
1729
1673
998
1695
1042
1898
1092
331
775
656
1452
1567
200
1265
1816
1606
455
697
657
1399
1526
214
1241
1441
833
111
1832
634
150
1861
399
106
561
790
993
1149
308
376
1981
411
494
349
983
249
720
1841
1299
392
1167
1872
127
1915
1473
434
1776
1643
1980
536
1670
1205
590
1336
1878
1068
1543
1238
725
1743
505
731
1929
333
922
272
1961
1490
279
976
436
874
1645
591
954
1198
316
863
1323
487
501
709
820
928
529
Dendogram – visualising the 'closeness' of variables
1487
1313
677
1630
414
478
1207
850
903
1854
1912
1109
14
1838
317
307
1203
1745
684
1139
1535
958
800
393
300
1675
486
347
242
1676
902
664
1576
1026
1973
1557
1423
535
1985
1173
532
516
675
104
33
1631
1463
93
1562
1846
154
1531
813
900
1291
226
306
1120
1819
1876
270
1114
258
62
1458
1888
1802
1.0
1378
1788
373
1381
689
51
203
779
550
102
1350
1235
921
1456
1651
700
1395
1050
805
741
871
5
1140
1717
827
481
1084
346
835
1206
1660
584
302
716
701
1848
593
969
1249
1656
933
714
1805
477
299
1843
1270
1712
1787
614
893
1445
1884
999
1809
1945
289
694
659
929
645
1433
1813
280
1903
597
353
1807
1393
1822
1283
605
846
301
1697
727
510
1085
113
726
1166
1498
327
1890
1716
636
1345
1768
1157
1694
1763
565
191
1955
945
864
1240
1556
994
1901
424
78
1436
1382
1880
1211
1154
1247
1881
915
713
332
1184
955
342
1100
600
129
830
456
243
761
1027
970
804
1771
1905
1325
1102
248
305
1071
463
57
1596
1798
241
633
1625
1605
691
1608
637
845
1734
1377
1862
852
19
46
1873
926
950
8
1964
1116
1870
1642
103
971
283
174
941
1142
1171
745
180
704
290
60
654
190
146
1678
1457
474
1967
686
170
1422
1405
1829
834
484
810
698
1992
987
199
1080
876
139
1667
155
273
338
829
972
1070
1252
1073
1740
1913
489
882
963
582
20
1794
496
26
1172
1994
539
1372
814
439
337
1208
819
391
132
521
435
1786
651
920
1462
1222
194
1001
1565
545
335
1579
425
979
1749
1213
444
1107
1747
558
1443
1936
269
1248
1554
752
770
1233
430
261
1409
554
504
1950
774
841
1688
1671
1290
1212
520
461
284
1065
931
875
1644
236
816
1314
277
366
1721
533
842
291
519
309
90
1736
1514
446
1282
966
429
1773
1495
1376
1864
1815
776
1127
187
596
360
1808
1481
498
1305
398
710
1078
1622
962
1351
1836
509
1106
1927
1273
1209
1900
1793
101
1373
1504
937
1276
1906
1624
255
135
65
1886
1284
511
253
1684
367
348
1699
978
229
1959
157
22
1129
1296
762
1257
1893
1303
186
1599
670
518
1021
1731
260
1761
930
56
1909
672
499
1523
1997
608
773
1455
1392
76
1636
1949
181
1855
1800
1126
1168
125
32
1366
1223
1943
924
1204
1775
1753
1885
851
832
1268
1750
12
1966
476
653
1315
1954
421
282
1367
1438
891
1730
1623
63
0.5
1144
1868
910
1148
1952
231
1294
579
564
858
1738
1932
823
1156
1544
849
1632
1910
116
1921
1907
87
1680
1349
1009
964
1578
1560
896
1403
328
500
Hierarchical clustering – top down or bottom up measure of client 'closeness'.
133
192
663
1412
1723
380
156
613
702
47
1993
973
410
160
625
72
1764
1465
389
1614
934
890
1724
1417
1031
182
1339
412
540
1450
840
655
938
453
15
1532
641
158
1540
1828
936
1250
862
869
121
796
769
1797
1825
706
1237
1176
1406
817
1344
1629
742
508
141
1
806
897
547
1510
173
475
1865
1352
1137
311
1418
594
492
1784
364
304
1506
1835
112
1264
1778
23
1002
240
599
61
756
825
503
1329
238
334
526
764
202
1263
1998
1097
1891
1748
1574
1334
1931
646
956
1615
1437
523
1302
1965
471
1842
682
848
730
1391
1918
838
1359
1939
143
1572
629
426
1600
1638
84
1419
1161
868
452
1801
450
620
108
580
538
1453
1244
247
1548
1330
1648
1134
1232
497
944
1052
592
85
688
916
91
1621
1682
1123
1791
1069
1803
256
404
953
2
Discriminant Coordinates audit-csv.txt
1558
1195
1633
1400
1341
1435
1739
965
59
1141
1484
899
1096
1371
185
175
215
801
566
1470
1923
469
1627
166
674
1439
740
3
1988
1338
577
35
1187
94
99
1928
413
9
621
854
961
1158
1951
416
1051
1700
140
468
1692
1821
1637
1503
1728
1034
734
406
17
1058
711
628
Variable Correlation Clusters audit-csv.txt
1769
1083
524
358
1355
381
722
980
206
919
1502
1182
86
1804
530
432
1496
1897
589
48
1262
144
534
1520
1568
808
1704
1183
843
757
712
38
1874
1703
925
986
1926
1879
583
1275
687
387
1849
1552
131
1088
402
438
1666
1512
1894
571
1863
1370
885
736
705
73
171
383
667
1476
718
542
1963
1272
939
908
1218
551
11
1214
1047
268
1292
1360
737
1006
1710
1404
1322
1609
791
123
802
647
28
1113
1857
948
1278
294
977
1407
1497
134
1477
1000
1564
293
1652
559
195
1413
1348
959
1082
1010
824
1258
1424
743
873
0.0
Age
Hours
Income
Adjus ted
Deductions
100
80
Performance (%)
60
Revenue
Adjustments
Strike Rate
40
28%
20
0
0 20 40 60 80 100
Caseload (%)
Rattle 2006-10-02 16:27:39 Stuart
Risk lift charts are a way of 'seeing' the improvement that an analytic model can produce and make the trade-off
between caseload and revenue more obvious to management.
In this example we can see that in this instance the Support Vector Model (ksvm) slightly outperforms the Random
Forest Model (rf) and the Logistic Regression model (glm) and these outperform the single Decision Tree (rpart)
which significantly outperforms the simpler Gradient Boost Model (gbm).
Where different models perform significantly better at different parts of the distribution, ensemble approaches can
be used to harness the right model for the right part of the distribution.
Case selection models need to be evaluated from a number of perspectives. Better selection models select more
true positives and true negatives and minimise false positives and false negatives.
Note that the strike rate metric [true positives/(true positives+false positives)] is like an iceberg - it is what you can
see and measure but it is only part of the picture as it does not provide details (except by inference) regarding
those not selected (false negatives and true negatives). Consider the following Taylor-Russell diagrams:
It would generally be serendipitous if an organisations case mix optimised its revenue yield. Analytic simulation
modelling techniques such as linear programming can provide insights to management regarding the optimum mix
of case types and their revenue and resource impacts. Such modelling can provide insights into key parameters
and their sensitivity to change.
A simple two case type example will be used to demonstrate how linear programming can be utilised in the
optimisation of case mix by a revenue authority.
Please read the following in conjunction with the graphs on the following two pages…
Assume that there are two case types that staff can work on – an Income Tax Case Audit and a GST Audit. For this
example let us say that an Income Tax Audit on average returned $500 in revenue and a GST Audit returned $200.
Assume that: we have three levels of staff involved in the casework – EL2's, APS6's and APS4's and that we have
five EL2's, four APS6's and four APS4's all of whom can work for 40 hours per week. An Income Tax Audit takes
on average two hours of an EL2's time plus two hours of an APS6 and one hour of an APS4. A GST Audit takes
on average an hour of an EL2's time plus an hour of an APS4. In this example we will say an EL2 costs $50 per
hour, an APS6 costs 20$ per hour and an APS 4 costs $15 per hour.
The linear optimisation question is what case mix optimises net revenues?
We can see that 2 Income Tax Audits = 5 GST Audits in terms of revenue. This is the slope of our revenue curve.
Graphically this produces a feasibility within which the actual case mix must sit. At the boundary of the feasibility
space an optimum case mix exists.
When we place our revenue curve onto the diagram and move it to its outer most point on the feasibility space we
identify our optimum case mix: 80 Income Tax Audits and 40 GST Audits. Our 'binding' constraints are the number
of effort hours available of our EL2's and APS4's. We can see at the optimum point we have spare APS6 capacity.
If we introduce a coverage constraint (must do at least 60 GST Audits) we can see that it reduces the feasibility
space and creates a new optimum point. The binding constraints at this point are the EL2 effort time and the GST
coverage constraint. We have spare APS6 and APS4 time.
This was of course a very simple example, however the methodology holds for greater numbers of case types and
resource types – it just can’t be shown graphically. (An example spreadsheet for more case types is included in
annex that uses the Excel ‘solver’ add-in.)
This type of analysis can be further enhanced by the modelling distributions rather than fixed amounts. (eg the
distribution of revenue, the distribution of effort time etc by case type. Software such as @Risk can be used for this
purpose. (See @Risk at http://www.palisade.com)
It is relatively easy for such staff to be deployed on research type projects to discover some interesting insight
about our client base. However this may not be the best method of prioritising tasks for the analytic capability.
When looking at where to apply analytic techniques to optimise revenue, a useful approach may be to consider
either revenue gains or staff savings by the relative strike rate being achieved. (Apriori, an area with a low strike
rate has the largest potential to improve. By looking at the revenue being collected or the staff utilised by 1-Sr, a
rectangle is produced whose relative size indicates a potential room for improvement.) In this example as large
business already have a relatively high strike rate and fewer staff there is smaller room for efficiency gains.
In this hypothetical example we can see that from an efficiency viewpoint the cash economy has a relatively higher
number of active compliance staff and a lower strike rate than large business – hence analytic enhancements to
the strike rate would provide a greater efficiency gain. From a revenue perspective it is less clear cut.
FTE
Cash Economy
Large
Business
1-Sr
Efficiency – Staff Focus
$
Large
Business
Cash Economy
1-Sr
Effectiveness – Revenue Focus
Where iterative refinement processes are used it becomes questionable as to what is a true false positive when
candidate cases are filtered out without a full review.
A richer picture emerges when one looks at the selection process end to end, taking into account the various
review and refinement approaches (and resources) used to ultimately allocate a case.
o The selection coverage rate (% of clients in the population run through the selection algorithm). Output is a
candidate case.
o The review coverage rate (candidate cases reviewed by an experienced officer prior to allocation).
o The data quality error rate (candidate cases not proceeded with because on review the data is incorrect).
o The selection error rate (candidate cases not allocated because on review they were not a strike).
We should note that the outcomes of the process are determined not only by the effectiveness of the case selection
process but also by the effectiveness of the treatment applied and the capability of the staff providing the treatment.
The case selection process can influence the effectiveness of the treatment by:
o selecting appropriate treatments based on attributes of the taxpayer (and historic effectiveness),
o narrowing the risks that need to be investigated (or highlighted in the case of passive treatments such as
letters); and
o supplying contextual information to direct the case worker to the likely source of non-compliance.
The change on resourcing requirements as strike rates change is dependent upon the relative trade-off between
the resources required for a productive case and those required for a non-productive case.
As non-productive cases typically are less resource intensive, it will not be a simple matter of say 10 fewer non
productive cases means that 10 more productive cases can be completed. Revenue projections made on the
basis of a 1 – 1 trade-off will inevitably be wrong in the majority of circumstances.
Revenue gains and/or staff saving can be derived from strike rate improvements. The impact of varying strike rates
upon the revenue authority is largely dependent on whether they are guided by revenue targeting or staffing
constraints.
This is not a straight trade-off of a non productive case for a productive case as often the effort time required for a
productive case will be significantly more than the effort to complete a non productive case.
At higher strike rates the same percentage strike rate change results in fewer additional productive cases than it
does at lower strike rates.
1500
1000
500
0
Sr
%
95
85
75
65
55
45
35
25
15
Strike Rate
The chart above plots the effects of a 3:1 effort time differential as strike rates change with fixed staffing on case
numbers.
With fixed staffing as the strike rate varies downwards the number of cases able to be completed grows non-
linearly in accordance with the effort time differential between a productive and non productive case and the
difference in strike rate: Srx/(1+[tPC/tNPC-1]*Srx)
If all cases are productive the number of cases able to be completed = Number of Staff (nS)/Productive Case Effort
Time (tPC).
If all cases are non productive then the maximum number of cases able to be completed (TCmax) = Number of
Staff (nS) /Non Productive Case Effort Time (tNPC).
For a particular strike rate (Sr) the number of productive cases (nPC) is given by:
nPC = TCmax*Sr/(1+[tPC/tNPC-1]*Sr)
(ie the change in number of productive cases times the revenue per productive case – assuming that revenue per productive case does not
change significantly as case numbers or the strike rate changes.)
5000
4000
3000
2000
1000
0
Sr 95% 85% 75% 65% 55% 45% 35% 25% 15%
Strike Rate
The chart above plots the effects of a 3:1 effort time differential as strike rate changes on case numbers with a
fixed revenue target and variable staffing.
With fixed revenue targets as the strike rate varies downwards the number of cases required to meet the revenue
target grows non-linearly as we must pump more cases through to maintain the revenue amount.
The number of productive cases required for the revenue is fixed irrespective of the strike rate. It is the total
number of cases needed to be completed, and hence the staffing, that is varied to maintain the revenue target.
In this situation if all cases are productive, the number of cases completed at this strike rate equals the revenue
target (R$) divided by the average productive case result ($rPC). (eg So a $3 million revenue target at an average
of $10,000 per case requires the completion of 300 cases.)
The number of productive cases is fixed at this level irrespective of the strike rate so: nPC = R$/$rPC
For a particular strike rate (Sr) the number of non productive cases (nNPC) is given by:
nNPC=(nPC-Sr*nPC)/Sr
(ie the time spent on productive cases plus the time spent on non productive cases)
(ie the change in number of non productive cases times the effort used on those cases - assuming that effort time per non productive case does
not change significantly as case numbers or strike rates change)
Subject matter experts use their experience to create Past cases of non-compliance are divided into
case selection rules to filter non-compliant clients out ‘successful’ cases (where relevant non-compliance was
from the broader population in respect of particular found) and ‘unsuccessful’ cases (where it wasn’t). The
risks. OLAP type approaches – slice n’ dice. data set is further divided into a ‘training set’ (used to
build the analytic case selection model) and a ‘validation
They focus on client features that they believe assist set’ to test that the model works.
in revealing whether a client is compliant or not.
Working with subject matter experts the analytics
These rules are often refined overtime to enhance modeller identifies client features that appear to be
the strike rate. associated with non-compliance. These are tested
statistically to see if they are associated with non-
The rules produced are generally subjectively compliance.
weighted to derive a client risk scores for work
prioritisation. The training data set is then essentially regressed
against the target set of successful cases to produce a
Eg (hypothetical example) risk scoring algorithm that optimises the probability of
o If client has WRE claim > $1,500 and Uniform > predicting a successful case from the data set.
$500 then risk score = 4
o If client has motor vehicle claim > $5,000 then The algorithm is then tested against the validation data
risk score = 2 set to see how it performs.
o If client has self education claim > $2,000 then
risk score = 3
o Add client risk scores to produce total risk score The rules produced are weighted by the algorithm to
o Select for review clients with total risk score > 8 produce a client risk score for work prioritisation:
If compliance staff skills, knowledge and experience are not up to the task then the treatment will not be optimal.
Staff cannot be overzealous, incompetent or corrupt in the performance of this work and continuous
improvement/quality assurance measures need to be in place to ensure that client treatment is optimally delivered.
o ‘Case call-over’ procedures so that all significant/old cases are reviewed by someone outside the team
o Back-log procedures that are triggered when the median age of the case pool reaches a particular
threshold above the standard.
Conclusions
The development of risk and intelligence approaches is reaching a new level of sophistication.
The convergence of computing power, machine learning approaches and statistical algorithms is providing revenue
authorities with the capability to be much more personalised in the targeting of treatments to clients.
The successful integration of these techniques with robust management approaches to the use of strategic
intelligence is enabling revenue authorities to be much more purposeful in how they treat the risks that challenge
them.
Harnessing this opportunity requires us to make the transition to a more quantitative evidence based approach to
management.
Stuart Hamilton
November 2006
9
Alan Kay. 2003 Turing Prize winner. Inventor of many of the computer GUI interface aspects we take for granted
today.
Clients A, B & C have paid the following tax over the last three years:-
A predictive algorithm estimates that the following tax may have been avoided or evaded (ie at risk):-
Client A Y3: 2,700, Client B Y3: 2,700 Client C Y3: 0
Y2: 2,700 & Y2: 0 Y2: 2,700
Y1: 2,700 Y1: 0 Y1: 2,700
where 1.5 = 8,100/5,400 & 0.474 = 8,100/(8,100+9000) & 1.21 = 0.474/0.393 for Client A
A single weighted 90/10 Revenue/Severity Multi Year Risk Score then gives:
Client A 1.47 [0.9*1.50+0.1*1.21] Client B 0.51 Client C 1.02
These can be transformed further into a score from 1.000 to 0.000 by setting the highest score equal to 1.000 and
scaling the other scores off this:
High score high risk so we select Client A, then Client C, then Client B…
Note this example did not use PV, weighting or discounting for clarity
It is important that our understanding of our clients move beyond simplistic data analysis approaches based on
summary descriptive statistics such as means, medians, modes and standard deviations.
In all of these cases a single point estimate based on the mean, median or mode will be significantly more incorrect
than an estimate produced by modelling the income using discriminatory variables. It is the difference between
trying to sum up a song with a single note versus having an MP3 version of it (compressed to the sounds that
matter and so still understandable).
We need to reach back into the distributions to understand the variance between client groups and find the
variables that matter (to answer the problem we are looking at.
With the computing power and algorithms now available we should be striving for the highest practical degree of
accuracy in our estimates. Anything less is almost bordering on professional negligence for an analyst.
1000
0 e+00
1500
Adjusted
80
All
0
40
0 500
1
0
0
Adjusted Adjusted
Rattle 2006-10-04 12:48:07 Stuart Rattle 2006-10-04 12:48:07 Stuart Rattle 2006-10-04 12:48:07 Stuart
1 1 1
0 200
Linear programming
Case Type
OptimisationV5.xls (...
case_work_calcs_v
2.xls (46 KB)...
@Risk from Palisade can be used with these spreadsheets to provide distributional input rather than fixed values.
This enables simulations such as linear programming to be more realistic in their projections.
See http://www.palisade.com/downloads/pdf/Palisade_RISK_0604PA.pdf
Review / Advise
Try to, but don’t Assist to
Educate / Market
always succeed comply
System changes
Simplification
Attitude to Compliance Prepopulation
Willing to do the
Compliance Strategy Make it easy
right thing
Another view of the end to end process from risk identification to case outcomes:
Ranked
Candidates
Coverage Operationalise
& Revenue Analytics
targets
Cases