Page 1
ALLSTATE INSURANCE GROUP
PRIVATE PASSENGER AUTO
COLORADO
The following addresses the concerns outlined by the letter received from the Division of Insurance
on 03/18/2016. The questions are restated below in bold font and immediately followed by our
response.

1. What specific characteristics are used to assign each policy to a micro-segment? In the
rate filing it states that a micro-segment is defined by the unique combination of the
following characteristics: territory, birthdate of oldest operator, years with prior carrier
and gender of oldest operator but in the phone call on February 2, 2016, Allstate stated that
these characteristics were arbitrary.
In discussing any specific structural component of Complementary Group Rating (CGR), it’s
helpful to think about the structure as a whole and what it’s designed to accomplish before
focusing on any individual aspect. CGR was designed to solve a problem that insurers have been
encountering more and more over time – how to incorporate new, different, but more predictive
loss models in an efficient manner.
For the purposes of this discussion the new loss model is not simply a refit wherein the same
variables, levels of variables, and structure as the existing loss model are used to develop
updated parameters/factors on a more recent data set. To fully illustrate the benefits of CGR, the
new loss model should be thought of as a much more substantial overhaul wherein there are
different variables (new or removed), levels of variables (more or less granularity), and/or
structure (how variables are considered, interactions of variables) than the existing loss model.
CGR enables Allstate to charge actuarially sound rates that incorporate the latest loss model
learnings in a much less disruptive and costly manner than prior methods.
Exhibit 1 contains a number of Figures that Allstate has provided to assist in the discussion of
CGR for this response. Figure 1 shows an example of an existing traditional rating plan – a
number of multiplicative rating steps, all of which have defined levels and factors for each
considered variable which may include interactions of variables. The traditional approach of
incorporating a new loss model is shown by moving over to Figure 2 – there are different
multiplicative rating steps (both in number and the steps themselves), all of which have defined
levels and factors which may be similar or different to the existing rating plan (for illustration
Step_2c shows that the interaction of variables is now different with different factors). In the
past, rating plan (loss model) overhauls have been done by simply putting the new rating plan in
place, either in a new underwriting company (which is not practical in the long run) or in an
existing underwriting company (which is disruptive to customers, and results in long delays
between loss cost estimate improvements).
Figure 3 shows an alternate approach where the new rating plan (loss model) can be incorporated
while keeping the existing rating plan from Figure 1. This is accomplished through adding a
granular adjustment factor to the end of the rating plan. The adjustment factor can be determined

 Page 2
on a granular enough basis to exactly replicate the traditional rating plan overhaul shown in
Figure 2. Said another way – the granular adjustment factor can be used in combination with the
existing rating plan to reproduce the premiums from any potential future loss model. This is an
extremely powerful concept that allows more frequent improvements to loss models that are no
longer confined to the structure of the prior model.
Logistically, if one wanted to charge the exact rating plan 2 premium, the granular adjustment
factor to arrive at a target premium is calculated by dividing the target premium by the premium
from the existing rating plan (shown in Figure 3 as Premium_2/Premium_1, which when applied
to the existing rating plan results in exactly the premium from implementing the traditional rating
plan overhaul). The main challenge of using one adjustment factor rating step (in combination
with the existing rating plan) to reproduce the premium from a different rating plan is – the
adjustment needs to be developed at a granular enough basis to capture all of the potential
differences in predicted loss from the two rating plans in consideration. There are dozens of
rating steps in use between the two rating plans, and each rating step can contain anywhere from
one to thousands of levels. The required adjustment factor to reproduce the new rating plan must
be granular enough to represent and capture all possible differences in every single rating
characteristic and associated factors between the two rating plans. While this will be discussed in
more detail in a later question, the existing sophistication of rating plans before CGR results in a
astronomically large number of risk characteristic combinations - so this level of granularity is
not new with CGR. The conclusion following this line of thought is that the only way to
accurately capture the differences between two rating plans for every possible situation is to
compare them at the policy level. If the granular adjustment doesn’t occur at the policy level,
then it’s not possible to reproduce the premiums from a new rating plan (loss model).
Since we’ve determined that the granular adjustment needs to take place at a policy level to fully
capture all the differences in dozens of characteristics and the associated factor predictions for
the levels within each, an approach can be developed that achieves the goal. The first part of
Figure 4 shows a structure where a granular adjustment (the adjustment factor) is assigned to
each policy number (the granular identifier that links the predictions between two rating plans or
loss models) and then used in the rating of the policy to produce the target premium. If policy
number is used as the linking identifier for the granular adjustment shown in Figure 3’s rating
plan, it’s possible to reproduce every single premium from the new rating plan in Figure 2 while
keeping the existing rating plan structure from Figure 1. This accomplishes the goal described at
the beginning of this response – incorporating a new, different, but more predictive loss model
efficiently. For discussion’s sake this hypothetical rating plan with a policy level granular
adjustment will be referred to as RP1_alt.
Up until this point the discussion has revolved around hypothetical rating plans to present many
of the key ideas of rating plan design with CGR. CGR accomplishes the exact concepts described
above and shown in the referenced Figures (RP1_alt), but has to do so with a few deviations that
while necessary because of technology constraints, unfortunately make the structure a little more
difficult to grasp.

 Page 3
Figure 4 outlines the differences in how CGR’s granular adjustment works in comparison to the
hypothetical adjustment from RP1_alt described above, but essentially accomplishes the same
thing.
In both cases there is an adjustment factor – RP1_alt has the Granular Adjustment while
CGR has the Complementary Group Factor.
o While the Granular Adjustment in RP1_alt could theoretically take on any
magnitude in value or number of decimal places, because of technology
constraints the Complementary Group Factor is restricted to 1000 possible values
that are established in 0.5% increments. This provides sufficient range to capture
the possible adjustment factor values, with enough delineation in values to
approximate any possible value extremely closely.
o CGR also has the Complementary Group which is just a three digit representation
of the factor value, and is used because of technology constraints. There is an
unchanging one-to-one relationship between the three digit representation and the
adjustment factor, so they can be thought of as equivalent information. In Figure 4
these components for both RP1_alt and CGR are accomplishing the exact same
result.
In both cases the adjustment factor is linked to a policy through some mechanism and is
then used in calculating the premium for that policy. The use of an adjustment factor with
the existing rating plan is key to the idea of efficiently incorporating new rating plans
(loss models).
In both cases there is a linking identifier between the adjustment factor and the rating of
the policy – RP1_alt has policy number while CGR has the micro-segment.
o Because of technology constraints, the policy number was not used as the linking
identifier in CGR. Instead, Allstate designed something that approximates the
effect of using policy number to achieve the level of granularity needed for the
goals described above.
o The micro-segment is defined as the unique combination of four characteristics –
territory, birthdate of the oldest operator, years with prior carrier, and gender of
the oldest operator. By evaluating these characteristics for each individual policy
(which is the sole method of assigning a policy to a micro-segment), the microsegment can be thought of as something effectively equivalent to policy number.
Rather than use policy number directly, the micro-segment approximates policy
number as an identifier.
In prior discussions with the Division, Allstate did express how the micro-segment variables
used were somewhat arbitrary. For clarification, the intended meaning of such statements were to
convey how the selected variables used to define the micro-segment (territory, birthdate of the
oldest operator, years with prior carrier, and gender of the oldest operator) were somewhat
arbitrary. For example, Allstate could have defined the micro-segment as a combination of the #
of operators on the policy, # of vehicles on the policy, model year of the oldest vehicle on the
policy, or any of the dozens of characteristics known for each policy. The purpose of the microsegment is to provide an identifier granular enough to approximate the policy number, which as
discussed above is necessary to fully reflect the differences between two rating plans. With a
given set of variables that define the micro-segment, the actual values for an individual policy
are used to relate the adjustment factor to the rating of a policy. The values of the selected

 Page 4
variables that define the micro-segment are not arbitrary, but have no impact other than to serve
as an identifier that links an assigned adjustment factor to the rating of a policy.
2. Every microsegment is assigned a Complement Group with a corresponding CGR within
the actuarially sound range. The CGR factors listed for Colorado range between 0.1066
and 9.3823. Please explain how a CGR factor is assigned to each policyholder when the
selection of CGR factors is so broad? Are policyholders with a similar risk profile assigned
the same CGR factor?
As described in the response to question #1, the CGR factor is used to capture the differences
between two rating plans (loss models) at a very granular level. When applied in combination
with the existing rating plan structure, the CGR factor results in an actuarially sound premium
that considers the new loss model. Figures 1-3 in Exhibit 1 demonstrate the concept of using a
granular adjustment factor with an existing rating plan to incorporate a new rating plan (or loss
model). Because of technology constraints, Allstate needed to define a set number of possible
values for the adjustment factor (CGR factor), which ranges from 0.1066 to 9.3823 in fixed
increments. Wanting to ensure that the factor could encompass the differences between the
existing rating plan and any potential future innovation in loss model development, Allstate
established a wide range beyond what is currently needed or used to incorporate the new loss
model. A histogram of the CGR factors used for existing Allstate Fire and Casualty Insurance
Company (AFCIC) insureds as of filing R28779 is shown below.

Use of CGR Factors in CO
Percent of AFCIC Policies

50%
45%

40%
35%
30%
25%
20%
15%
10%

5%
0%
0.1066 to
0.25

0.25 to
0.50

0.50 to
0.75

0.75 to
1.00

1.00 to
1.25

1.25 to
1.50

1.50 to
2.00

2.00 to
3.00

3.00 to
9.3823

CGR Factor Range

It can be observed from this histogram that while there is a wide range of CGR factors available
for use in the CGR structure, a relatively small portion is actually being used. There are very
small amounts of policies in the CGR factor buckets further from 1.00, and none in the greater
than 3.00 bucket.

 Page 5
In discussing how the CGR factor is assigned for a given micro-segment (which again serves as a
structural approximation for policy number), it’s important to remember what the CGR factor is
designed to encompass. The CGR factor is used to capture the differences in two actuarially
sound estimates of rate. Those differences are the result of the loss characteristics for a given
policy and what the two different rating plans (loss models) say about how those characteristics
affect the estimated loss. The CGR factor represents differences in loss estimates (which are
driven by the effect of all loss characteristics), not any specific grouping of loss characteristics.
Figure 5 in Exhibit 1, carrying forward the example from Figures 1-4, outlines how the concept
of two actuarially sound estimates of rate is translated into a selected CGR factor. The first
section shows two actuarially sound estimates, one from the existing rating plan (Premium_1)
and one from the new rating plan (or loss model, Premium_2). Since both Premium_1 and
Premium_2 are actuarially sound estimates, it follows widely accepted actuarially practices that a
selection between those two estimates is also actuarially sound. The range between these two
estimates form the boundaries of what we define as the actuarially sound range of premium.
The bounds of the actuarially sound range can also be thought of in terms of the CGR factor
(adjustment factor) rather than premiums. CGR_Factor_1 bounds one side of the range, and
CGR_Factor_2 bounds the other side. Any CGR factor between those two actuarially sound
estimates is defined as actuarially sound. The width of this range is dependent on the loss
characteristics of the policy, how those characteristics translate to estimated loss for each of the
rating plans (or loss models) being considered, and the difference in the estimates.
Given the complexity and granularity of both the existing rating plan and the new loss model, it’s
very difficult to make any generalizations that certain characteristics are associated with a certain
difference in loss estimates (in direction or magnitude). With dozens of rating steps in
calculating each loss estimate, individual rating steps containing up to thousands of different
levels and associated factors, there are too many influences on the estimates (and differences
between them) to make any direct associations to specific characteristics. Two policies with
similar differences in loss estimates do not necessarily have any shared characteristics – the only
thing that can be said for sure is that their characteristics and resulting loss estimates produced
similar difference in loss estimates from two separate class based analysis.
Unless the two loss estimates (Premium_1 and Premium_2) are very close to each other or
exactly the same, there can be multiple CGR factors within the actuarially sound range. Allstate
uses an objective method in determining which of the actuarially sound CGR factors is selected
for each micro-segment:
The selected CGR factor is within the actuarially sound range for each micro-segment.
This ensures that the rate being charged for each micro-segment is appropriate and based
on two actuarially sound loss estimates from class based analysis.
The resulting percent impact as a result of applying the selected CGR factor for each
micro-segment is within a defined range. Depending on the width of the actuarially sound
range, this can further restrict the range of possible CGR factor selections.
The overall rate level of the entire book is maintained. This ensures that applying the
CGR factor to each micro-segment does not change the overall amount of premium (rate

 Page 6
level) for the book. Any such changes will always be filed with the division with
accompanying rate level indication support.
The overall estimated retention of the book is maximized. The estimated retention is
calculated as the expected number of vehicles that will retain through the upcoming
policy period, and is estimated using two class based retention models that have been
discussed previously with the Division.
To recap:
The range of possible CGR factors needed to be established for technology purposes, but
only a narrow subset of the possible values have been used to date.
Two actuarially sound class based estimates of loss are used to define the actuarially
sound range of CGR factors.
The width of the range of actuarially sound CGR factors is caused exclusively by the loss
characteristics and class based factor predictions from the respective rating plans (loss
models).
The assignment of a CGR factor to a micro-segment is based solely on an objective
analysis that can only result in an actuarially sound premium that also maximizes the
retention of the whole book of business.
3. Allstate provided the selected and indicated rate relativity for each Complementary
Group. Please explain why some Complementary Groups receive full indication while some
Complementary Groups show an indicated rate relativity change over 100% and have a
selected rate relativity change of 0%? How will the policyholders in these Complementary
Groups ever reach full indication?
While the terminology used in describing elements of CGR can understandably be difficult to
grasp, Allstate feels it is important to make some clarifications to ensure that all parties are
correctly interpreting the provided information (for prior communications as well as this one). To
clarify, Allstate has in the past provided the indicated and selected percent changes for each
micro-segment. As the micro-segment is an approximation of policy number and representative
of the loss estimate from the complete set of class based characteristics, it is an appropriate
comparison to make. Allstate has not provided a similar view for each Complementary Group, as
they are merely a three digit representation of the CGR factor. Complementary Groups are not
classes or identifiers representing a set of loss characteristics (such as policy number or microsegment), and should not be analyzed as such.
The actuarial profession and insurers have continuously sought to improve predictions of loss
through the use of more advanced analytical techniques, incorporating more detailed data, and
increased computing power that makes it possible. While standard practice when predicting loss
for personal auto insurance continues to utilize class based analysis, the end result of applying a
class based model is an estimate for an individual risk (policy or vehicle). With any new loss
model, there are significant differences in the estimates of loss for individual risks compared to
the prior model. This can be attributed to differences in modeling data, differences in analytical
techniques, substantial effects from new variables/interactions/classification granularity, the
cumulative effect of many smaller differences predictive effects, and more than likely a
combination of all of the above. It is extremely common for the differences between loss models

 Page 7
to resemble something that looks like a bell curve. For example, a histogram of the indicated
CGR factors (which fully captures the differences between the existing rating plan and the new
loss model) for existing Allstate Fire and Casualty Insurance Company (AFCIC) insureds as of
filing R28779 is shown below.

Indicated CGR Factors in CO
Percent of AFCIC Policies

50%
45%

40%
35%
30%
25%
20%
15%
10%

5%
0%
0.1066 to
0.25

0.25 to
0.50

0.50 to
0.75

0.75 to
1.00

1.00 to
1.25

1.25 to
1.50

1.50 to
2.00

2.00 to
3.00

3.00 to
9.3823

CGR Factor Range

Again while this histogram shows the indicated CGR factor between the existing rating plan and
the new loss model, it truly is representative of the typical differences from any new loss model
developed in remotely recent times. Insurers used to struggle with the disruption caused from
these differences in loss estimates once every 5-10 years (the frequency of building and
implementing new loss models). While there are likely many causes (healthy and competitive
market for auto insurance, advancements in analytical techniques, computing power, innovation,
etc), there has been a trend towards more frequent and substantial revisions in loss models. It’s
not unrealistic that the timeframe of building and using new loss models could be happening
once every 1-3 years going forward, which further exacerbates the problem of directly
implementing loss estimates that are substantially different from the prior estimate.
An insurer’s prediction of loss for a given risk, while a best estimate and continually improving,
are ultimately just estimates that will change over time. Directly implementing a new loss model
whenever it is developed would be extremely disruptive, and could likely result in policyholders
experiencing oscillating rate changes of different magnitudes and direction every few years from
solely a revision in loss estimate that will continue to change over time. As a result, insurers
typically rely on making judgmental, subjective factor selections when incorporating a new loss
model estimate, rather than charging exactly or completely reaching the new estimate since it
will likely change again. CGR provides a significant improvement over traditional factor
selection as described in this response.
Many of these ideas and goals have been built into the design of CGR and the process used to
assign CGR factors (granular adjustment factors) to micro-segments (approximation of policy

 Page 8
level identifier). The idea of CGR and using a granular adjustment factor is the direct result of
recognizing the increased frequency of loss model advancements along with the significant
implementation costs and disruptions that follow. This structure allows for the direct
implementation of the updated loss estimates from a new loss model, or implementing an
actuarially sound rate that uses the updated loss estimate to as a key piece of information. As
described in the response to question #2, the process for assigning a CGR factor to each microsegment is:
The selected CGR factor for each micro-segment is within the actuarially sound range –
defined as those between the current rating plan and the new loss model. This addresses
the goal of incorporating an updated loss estimate from a new class based loss model as
a more accurate prediction of loss.
The resulting percent impact as a result of applying the selected CGR factor for each
micro-segment is within a defined range – which depending on the width of the
actuarially sound range can further restrict the range of possible CGR factor selections.
This addresses the goal of limiting policy disruption as the result of an updated loss
estimate, since the loss estimate will likely be revised again and potentially substantially
different within a few years.
The overall rate level for the entire book is maintained. While this doesn’t connect
directly to the goals described above, it does ensure that the analysis and changes
between overall book rate level and more granular risk classification rates are kept
separate.
The overall estimated retention of the book is maximized as estimated using class based
retention models. This criteria is an objective implementation of what has been done
subjectively in the past with factor selection to address policy disruption and changing
loss estimates.
All CGR factors are assigned to micro-segments using exclusively these criteria. Examining the
selected and indicated percent changes for any individual micro-segment is explained within
these criteria. The range of CGR factor options for any individual micro-segment is the overlap
of the actuarially sound range and defined percent impact range. Once the range of CGR factor
options is established, the actual selection is based on two things:
1. Ensuring that the selection of CGR factors for all micro-segments result in the correct
overall rate level.
2. Simultaneously considering the estimated retention for all micro-segments for all CGR
factor options.
There is no mechanism in this process to prioritize or force movement of micro-segments to fully
indicated or any specified magnitude. CGR is not designed to immediately or over a specified
amount of time move the rate for policyholders to the updated estimate of loss – it’s designed to
efficiently incorporate the new loss estimate as a key piece of information in offering updated
actuarially sound rates. Just as with traditional factor selections, this can result in different
policyholders being moved by varying degrees (or not at all) towards indicated. Exhibit 2
contains a hypothetical example of revising class level factors towards an updated view of
indicated by class, and the resulting range of impacts (in both magnitude and direction) relative
to indicated. This example has been shared with the Division before (conference call on
3/25/2015 with a copy provided electronically on 3/31/2015), but is included here for reference.

 Page 9
Please note that CGR represents a major improvement over the traditional factor selection
approach shown in this example, as while factor selection results in policy premiums moving in
the opposite direction of the indicated loss, CGR does not.
4. In the Colorado filing, it states that one requirement of CGR is that for each microsegment, the resulting impact must be within a set range. In this filing, each microsegment’s impact must be within -20% and +20%. Does Allstate consider this a form of
rate capping? If so, how does this rate capping comply with Bulletin B-5.32?
Allstate is charging the filed manual rate for all customers, and is not using any capping
mechanism that would vary premiums charged from the manual rate. Allstate does not consider
the percent impact criteria of assigning CGR factors to micro-segments to be a form of rate
capping. It has been considered a widely accepted practice to incorporate judgment and business
considerations when making selections for rates – whether how far to move towards an aggregate
rate level, how far to move towards an updated factor estimate for the same loss model structure,
or how far to move a collection of factor changes to limit the combined impact to limit
policyholder disruption. The use of the percent impact and retention criteria discussed above are
merely objective implementations of the considerations that have been historically considered
and implemented subjectively on a case by case basis.
The selected CGR factor for each micro-segment meets the criteria discussed in the prior
responses as of a data evaluation date, and is provided to the division in a rate filing. The
assigned CGR factors are fixed and unchanging unless Allstate submits a filing with revised
assignments. The rates for policyholders do not automatically move at each renewal or any other
time period. Allstate performs a loss cost based analysis wherein an individual CGR factor is
selected, filed, and charged just like any other rating plan factor selection.
5. For the issue of credibility, Allstate stated that the new loss and retention models are
class based and therefore credible. How are the Complementary Groups considered
credible if there are some Complementary Groups with only one or a few policyholders?
Many years ago, the rates offered to customers through early versions of classification
ratemaking were essentially a table of a rates for a combination of a small number of
characteristics. For example – risks within a certain age range and a certain # of accidents in the
prior 3 years would be offered a rate. Classification ratemaking evolved into identifying credible
levels of characteristics that have a consistent, predictive, multiplicative effect on losses. Modern
rating plans have identified dozens of characteristics with up to thousands of levels within a
given characteristic, all used to produce an estimate of loss. To accompany this discussion, actual
information about the rating plan in place in CO Allstate Fire & Casualty Insurance Company
(AFCIC) excluding the introduction of CGR will be referenced.
AFCIC’s existing rating plan contains 43 multiplicative rating steps, with each step ranging from
one to thousands of levels. The basis of this rating plan was a credible class based loss model
built on multiple years of countrywide policy experience - a GLM that fit parameters (factors)
for each specified characteristic levels to as closely as possible reproduce the actual losses
experienced. There is a substantial amount of effort in building, testing, and validating GLMs

 Page 10
that are beyond the scope of this response. But the end result is a credible specified factor for
different levels of characteristics that when combined multiplicatively, produce credible loss
estimates for that unique combination of characteristics. This is a very important concept in
interpreting class based loss models (rating plans) – while the factors are fit for each specified
class (levels defined for a given characteristic), the output is an estimate for the cumulative effect
of all applicable loss characteristics. Loss models do not fit a factor to capture the loss potential
of an individual risk, but by fitting many characteristic levels that each capture a predictive
multiplicative effect, there is a unique loss estimate for any given set of loss characteristics.
It’s easy to underestimate the sophistication and number of price points produced from a
traditional class based rating plan. Continuing the example of the existing AFCIC rating plan:
There are 43 multiplicative rating steps (excluding CGR).
The rating steps range from one level (the “Rate Adjustment Factor” which is used as an
efficient way of increasing all base rates), to tens of levels (the “Safe Driving Club
Factor” which varies based on the # of operators that meet an incident criteria compared
to the total # of operators), to thousands of levels (the steps which reflect the amount of
insurance purchased for different coverages).
Ignoring some of the more complex steps of the rating plan (geometric averaging, vehicle
level characteristics, etc.), the combined effect of these 43 multiplicative rating steps
results in over 6*1032 combinations of rating characteristics in the rating plan. This
approximation significantly underrepresents the total number of combinations because of
the complexity of several excluded rating steps from this calculation.
In building the loss model used for the existing rating plan (prior to CGR), Allstate did not
develop factors or try to predict the losses for 6*1032 types of risk. But this level of granularity is
the result of any class based loss model (rating plan) with any degree of modern sophistication –
number of rating steps and number of defined levels within each.
In examining its rating plans and loss models, Allstate has yet to find two policies in any given
state (including CO) that share a complete set of risk characteristics. A difference in any of the
risk characteristics between two policies produces a different definition of risk and a different
estimate of loss. Given the number of rating steps, the number of defined levels within each, and
the number of potential policies in a book of business, with all of the provided information it’s
relatively easy to see how there are no two identical risks (policies with the same risk
characteristics). While more sophistication in a risk classification system (rating plan or loss
model) makes the likelihood of there being two identical risks even more remote, it’s still very
unlikely to find two identical risks through the lens of what would be considered an extremely
simple rating plan by today’s standards. For example – using only a handful of the existing rating
plan steps result in a unique identification of risk characteristics for each policy in this CO book
of business. Any class based model estimate is not developed at the individual policy level, but
the effect of using multiple classes in developing and scoring the model result in unique loss
estimates for unique combinations of risk characteristics.
One can use the number of rating steps as an approximation of rating sophistication and
granularity for illustrative purposes, although it is not perfect since not all rating steps are equal
in granularity. A cursory look at recent filings of the top personal auto insurers in CO show that

 Page 11
Allstate’s primary competitors have rating plans that are primarily in the 40-50 step range, with
the lowest being 24 and highest being 60. Again this isn’t a perfect representation of
sophistication because of the varying granularity of individual rating steps (ranging from simpler
than Allstate’s approach to dramatically more complex in some cases), but it does establish that
prominent insurers are all in a somewhat similar position – an astronomically large number of
risk characteristic combinations, and likely no two identical risks.
In personal auto insurance, the prediction of loss for a given risk is generally derived from the
combined effect of many class based effects, but the prediction is unique to that complete set of
risk characteristics. The responses to the prior questions discuss the significant differences
between loss models in both how they are built (characteristics, levels of characteristics,
interactions) and the resulting predictions (the effects of how it is built, factor predictions that
produce total loss estimates). Just as it has become increasingly difficult to think of risk as an
individual characteristic level (since there are dozens of characteristics that define a risk), it has
also become difficult to compare different loss models (rating plans) at an individual
characteristic level (since there are dozens of characteristics that define a risk in different
manners between the loss models).
It’s been established in this discussion that the existing widely accepted class based rating plans
(loss models) produce estimates for an astronomically large number of risks (unique risk
characteristic combinations). It follows that the only way to fully capture the differences between
the two rating plans (loss models) is to either implement the rating plan structure directly, or use
a very granular adjustment factor. As discussed in the response to question #1, all of the structure
and terminology associated with CGR is designed to enable a granular adjustment factor to
incorporate the updated loss estimate from a new class based loss model within the existing
rating plan. The new loss model that defines the updated loss estimate (which will be discussed
in more detail in a later question) is conceptually the same as the existing rating plan in that - it is
a credible class based loss model built on multiple years of countrywide policy experience, but
different in the manner that any new loss model is different through structure and predictive
effects.
In assigning the CGR factor for an individual micro-segment (approximation for policy), two
credible and actuarially sound loss estimates are used to define the actuarially sound range. The
two estimates are calculated from two credible class based loss models (rating plans) that capture
different predictive multiplicative effects from credible characteristic levels, but as discussed in
detail the resulting estimate is unique for a unique combination of risk characteristics. Again,
Allstate did not develop loss models to fit a factor to an individual risk, but the result of a class
based model with any degree of sophistication results in a credible prediction that is unique for
each individual risk. An objective process then selects a rate between these two credible, class
based loss estimates.
Micro-segments, Complementary Groups, and CGR factors are not classes and were not used as
a part of any loss model to calculate or develop estimates of loss. As discussed throughout this
response, these are structural elements of CGR that allow the selection of an actuarially sound
rate for a risk based on two credible actuarially sound estimates of loss from class based analysis.

 Page 12
6. Some factors in the retention model are not loss or expense based. Are these factors and
other retention adjustments applied at an individual policy level?
With any class based model (loss or retention), large datasets are used to identify predictive
characteristic levels and fit parameters (factors) to capture the effects that in total most closely
match actual experience. After the development phase, class based models are then used by
taking individual policies and scoring (running through the already built model) their individual
characteristics to obtain a predicted estimate. While class based models are designed to capture
predictive proportional effects associated with the specified classes, the results are always a
prediction for the individual policy. With this understanding, every characteristic from every
class based model is applied at the individual policy level, although the predictive effects are
built using a credible class based analysis.
One of the advantages of using a granular adjustment factor to incorporate a new loss model is
the flexibility and efficiency it provides going forward. When Allstate develops a newer loss
model with more accurate estimates of loss, the CGR process can be updated relatively easily to
use that model for the source of the indicated rate (one of the two actuarially sound estimates that
define the actuarially sound range). This is in stark contrast to the traditional process of
implementing a new loss model through the use of a new underwriting company or struggling
with the significant complications and disruption from implementing it directly in an existing
book of business.
In addition to the flexibility in incorporating new loss estimates, the structural design of CGR
also allows for flexibility in many of the other criteria used in the process of assigning CGR
factors – including the retention model. A modified or new class based retention model can be
incorporated in the process of assigning CGR factors. In fact – Allstate is currently in the process
of developing a new class based retention model that will be used developing future CGR factor
assignments in CO AFCIC. This new retention model is being fit on more recent data, and is
designed to only include classes related to loss, expense, or premium change.

7. Please provide a list of what states:
Filed/Approved the Allstate filing with no issues
Have ongoing questions/discussions with Allstate about the filing
Requested Allstate to revise the filing prior to it being filed/approved
Allstate has withdrawn the filing
Allstate has not made a filing
States filed/approved as-is (no issues or no active/recent discussions)
Arizona
Arkansas
Delaware
Idaho
Illinois
Iowa
Kansas

 Page 13
Michigan
Missouri
Nebraska
New Hampshire
Oklahoma
Oregon
Tennessee
Texas
Utah
West Virginia
Wisconsin
Wyoming
States with ongoing questions/discussions
Colorado
Maryland
Minnesota
Ohio
States that requested revisions (before or after filing/approval)
Indiana
Kentucky
Montana
Nevada
New Mexico
South Carolina
Virginia
States withdrawn
Alabama
Florida
Georgia
Louisiana
Maine
Pennsylvania
Rhode Island
States not filed
Alaska
California
Connecticut
District of Columbia
Hawaii
Massachusetts
Mississippi

 Page 14
New Jersey
New York
North Carolina
North Dakota
South Dakota
Vermont
Washington

8. In the instances that Allstate has revised a filing, please describe what sections or factors
were removed from the filing.
In states where a Department has requested revisions, Allstate has modified CGR in the
following ways:
Impose a minimum amount of movement towards indicated potentially beyond what
would have been selected from the objective assignment process.
Remove the retention models as an input in the assignment process to either use CGR as
a means to move rates towards indicated in a uniform manner.
Remove the new loss model and retention models as inputs, and simply uniformly limit
the rate impact of rating factor changes.
Modify assignments to limit policyholder impacts.
Remove variables from the existing retention model not related to loss, expense, or
premium change.
Override some loss model indicated factors with judgmental selections.
Modify the assignment for micro-segments with no existing Allstate Fire and Casualty
Insurance Company policyholders.
Modify assignments so impacts result in the exact (or less than) targeted percent change
rather than closest to the targeted percent change.
9. For Allstate’s most recent filing, please provide the following information relating to
Allstate’s Colorado Private Passenger book of business:
a. The total number of insurance policies.
As of the most recent Complementary Group assignment table update filing (R28779 which went
effective on 10/3/2015), there were 112,721 policyholders in Allstate Fire and Casualty
Insurance Company (AFCIC).
b. The number of classes after the class based loss model is applied. For each class, please
list the indicated rate and the number of insurance policies in each class.
The estimated loss from a rating plan or loss model is the product of a number of multiplicative
rating steps, with each rating step contributing proportional predictive effects for risk
classification levels that in combination with all other rating steps produce a rate (loss estimate)
for an individual risk. Any view of risk (premium, rate, loss estimate, listing of characteristics)

 Page 15
cannot be fully described by an individual class (risk classification level), but the combination of
all applicable classes.
Allstate’s existing rating plan before the application of CGR contains:
43 multiplicative rating steps.
With a simplistic view (excluding some of the more complex rating steps) of the rating
plan, there are over 1,400 classes (classification levels within the rating steps).
Continuing with a simplistic view, the combined effect of the rating steps and
classification levels within each result in over 6*1032 combinations of rating
characteristics.
Allstate’s new loss model used to calculate the indicated rate in the CGR assignment process
contains:
70 rating variables used across 114 multiplicative rating steps.
Continuing with a simplistic view, there are over 1,800 classes (classification levels
within the rating steps).
Continuing with a simplistic view, the combined effect of the rating steps and
classification levels within each result in over 4*1072 combinations of rating
characteristics.
Allstate is committed to assisting the Division in their review of CGR and providing the
information useful in that regard. When reviewing rates, Allstate does not believe that the
Division will gain any value from reviewing average rates on a univariate, class-by-class basis.
This is because any rate is determined (and developed) from the full combination of classes that
a policy belongs to, and any single class that a policy belongs to is only one small piece of how
their rate is determined. Said another way - a univariate class-by-class view of average rates
does not reflect the multivariate nature of how risk is assessed through class based rating plan
development. We are therefore providing what we feel is truly useful information to the Division
in this reply, rather than a creating a view that would be extremely time consuming and
inefficient for both Allstate to produce and the Division to review.
Exhibit 3 contains for each risk – which is defined as the unique combination of risk
characteristics – the indicated rate as calculated from the new class based loss model as of the
most recent Complementary Group assignment table filing (R28779). Rather than list out the
complete set of risk characteristics for every combination that exist in the CO AFCIC book of
business, an identifier for each policy is listed as representation of all the risk characteristics.
While the identifier is not the micro-segment, it is similar in concept and equivalent to a listing
of policy numbers.
Since the Division is requesting data on the indicated rates themselves, we believe Exhibit 3 is
the most useful way for Allstate to provide the information to the Division. It fully captures the
predictions of the new loss model for each type of risk (complete set of risk characteristics)
without the significant issues that come with evaluating rates as an average on a univariate classby-class basis. However to assist the Division in its review Allstate is providing a more
summarized view that is described in the response to question #9.c.

 Page 16
If the Division wants more information on the classes utilized in new class-based loss model, a
view of the indicated rating factors (as opposed to average rates) would be the most useful view
for the Division – we would be happy to provide this information at the Divisions request or any
other information the Division is interested in such as descriptions of the modeling process,
goodness of fit views and metrics for variables or in aggregate, or a complete listing of all classes
and corresponding parameters (factors) used in calculating the predicted estimates.
c. For each class, please list the current and proposed rate for each insurance policy in the
class after CGR and the retention model are applied.
Continuing the discussion in the response to question #9.b and the data provided in connection to
filing R28779, the current rate for each risk has been provided in Exhibit 3. Please note that the
current rate for each risk is defined as the rate (combination of existing rating plan and assigned
CGR factor) prior to their renewal onto the updated assignment table going into effect
(10/3/2015), and as discussed in that filing adjusted for overall changes to the rate level where
applicable.
The proposed rate for each risk has also been provided in Exhibit 3, which is the result of the
existing rating plan and assigned CGR factor as of that filing. As discussed in detail throughout
this response, the goal of all the complexity underlying CGR is to efficiently incorporate a new
estimate of loss in offering an actuarially sound rate that resides within the actuarially sound
range of rates – defined between current and indicated. The proposed rates in Exhibit 3 are
consistent with this goal and the described structural approach of CGR. The following are
additional pieces of information that may be helpful in evaluating Exhibit 3:
The selected rate resides within the actuarially sound range (current and indicated) for
nearly all risks, other than cases where the differences are very minor (due to 0.5%
increments in the available CGR factors and rounding that occurs) and an idea discussed
below. This is only possible with a granular adjustment factor or the direct
implementation of the new loss model.
As discussed in filing R28779, the selected CGR factor (and as a result the selected rate)
for micro-segments occupied by policies that renewed during the time the two prior
Complementary Group assignment tables were in effect were not changed. This results in
a larger amount of risks receiving no selected change from current than in a typical
assignment table change – a decision based on the magnitude of the changes resulting
from those tables.
The use of an adjustment factor at a very granular level is key to the rating plan concepts
described throughout this response and in Exhibit 1 Figures 1-5 and implemented with
CGR. The micro-segment is used as the linking identifier (structural link between the
new class based loss model and the rating of a policy) as an approximation of the effect
of policy number, which is not used because of technology constraints. While the microsegment and policy number provide equivalent granularity for the overwhelming majority
of cases, there are a small number of situations where the micro-segment is not unique to
an individual policy number. In those cases the actuarially sound range is satisfied at the
micro-segment level, which may result in different impacts for the individual policies that
reside in a given micro-segment.

 Page 17
As described in responses to questions received from prior filings (2/24/2015 and
4/8/2015) there are a small number of cases where Allstate was, because of technology
issues related to bringing together all the necessary detailed information, unable to
calculate an indicated value from the new loss model for an individual risk. In those few
cases where an indicated rate couldn’t be calculated, no change from the current rate was
selected or listed in Exhibit 2.
With the application of the CGR rating step, Allstate is offering an actuarially sound rate that in
an efficient manner uses the estimate from the new loss model along with objectively applied
retention considerations. There is no explicit factor or adjustment from the retention model, as
the inputs in the objective CGR factor selection process are not separated or applied individually.
These inputs, discussed in much greater detail throughout this response, are simultaneously
reflected in the selection and application of one factor that results in a reasonable and actuarially
sound rate.
While the detail included in Exhibit 3 is the most accurate manner to review our rates, we wanted
to provide the below summarized view as well to assist the Division in their review. The
indicated CGR factor, which is one of the two actuarially sound estimates that define the
actuarially sound range, fully captures every single difference between the existing rating plan
and the new loss model (including the variables used, levels within variables, interactions of
variables, modeling approach, and fitted factors). Since the levels of a univariate view of the
indicated CGR factor reflect the multivariate effects that go into predicting both actuarially
sound estimates (the existing rating plan and the new loss model), it is a more representative
(although not perfect) view of risk than what univariate views on a class by class basis would
provide.
Average Premiums
by Indicated CGR Factor Vigintile (Ranked 5% Groups)
1400
1300

Average Premium

1200
1100
1000
900
800
700
600
500
400
1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

Indicated CGR Factor Vigintile (20 Percentile Ranked Groups)
Average Current

Average Selected

Average Indicated

Average Excluding CGR

The indicated CGR factor, which is when applied to the existing rating plan in exactly the
indicated rate from the new loss model, was ranked and divided into twenty evenly sized groups

 Page 18
of policies (vigintiles). The above graph then shows the following which are consistent with the
goal and the described approach of CGR:
The indicated average rate, as calculated from the new class based loss model, increases
from left to right as the indicated CGR factor increases. This demonstrates that as
expected the multivariate effect captured in the indicated CGR factor is directly related to
the indicated rate.
There are significant differences between the average indicated from the new loss model
and the average rate from the existing rating plan excluding the application of the CGR
factor (as if the CGR rating step did not exist).
Compared to the rate from the existing rating plan excluding the CGR factor, the average
current rate (prior to the updated assignment table going into effect) is closer to the
indicated rate from the new loss model. This demonstrates that the existing rates charged
before the referenced filing (the latest assignment table update) are closer to indicated
than if the CGR factor was never used.
The average selected rate from the referenced filing (the latest assignment table update
going into effect) is even closer to the average indicated than the average current rate.
This demonstrates that assignment table updates result in actuarially sound rates that are
closer to indicated than what is currently in place.

 Exhibit 1

Figure 1 - Traditional Rating Plan

*
*
*
*
*
=

Rating Plan 1
Step_1a
Step_1b
Step_1c
…
Step_1r
Step_1s
Premium_1

Figure 2 - Traditional Rating Plan Overhaul

# of
Accidents
1
2
3
1
2
…

# of
Drivers
1
1
1
2
2
…

Step_1c
Factor
1
1.02
1.05
1.04
1.15
…

Traditional method of incorporating
a new loss model. Different
structure, variables, and factors
replacing the existing rating plan.

Rating Plan 2
Step_2a
Step_2b
Step_2c
…
Step_2y
Step_2z
Premium_2

*
*
*
*
*
=

# of
Accidents
1
2
3
1
2
…

Age of Youngest
Driver
17
17
17
18
18
…

Step_2c
Factor
1.2
1
0.95
1.17
0.98
…

Alternate approach to incorporate Rating Plan 2
while keeping the existing structure of Rating Plan
1. Significant cost savings as rating plan
sophistication increases more frequently.

Figure 3 - Alternative Rating Plan "Overhaul"

*
*
*
*
*
*
=

Rating Plan 1
Step_1a
Step_1b
Step_1c
…
Step_1r
Step_1s
Granular
Adjustment
Premium_3

# of
Accidents
1
2
3
1

# of
Drivers

Step_1c
Factor

1
1
1
2

1
1.02
1.05
1.04

2

2

1.15

…

…

…

If Granular Adjustment =

Premium_2
-

Premium_1

for each policy, then the resulting Premium_3 = Premium_2

Premium_3 = Granular Adjustment * Premium_1 =

Premium_2
-

Premium_1

* Premium_1 = Premium_2

Regardless of Rating Plan 2's structure, variables, or factors, the resulting premium can be achieved by
keeping the existing rate plan (Rating Plan 1) and an Adjustment. The Adjustment must be very granular to
capture the differences in Rating Plan 1 vs Rating Plan 2 for every unique combination of all risk
characteristics (Step_1a-Step_1s and Step_2a-Step_2z). Due to the sophistication present in each of the
rating plans, the required granularity to address every possible risk is policy level.

Figure 4 - Structure/Implementation of the Granular Adjustment
Figure 3 Alternative Rating Plan "Overhaul" Structure
Rating of
Policy

Policy Number
(Linking Identifier)

Granular Adjustment
(Adjustment Factor)

Although some of the interim steps of the structure are different in
Complementary Group Rating, it essentially achieves the same effect as
the structure outline in Figure 3 - linking a Granular Adjustment Factor to
the policy for the purpose of incorporating a new rating plan while
keeping the existing structure.

Complementary Group Rating Structure
Rating of
Policy

Micro-segment
(Linking Identifier)

Due to technology constraints, the micro-segment is used
rather than policy number. By using the unique combination of
four selected characteristics, the micro-segment approximates
the effect of using policy number as a granular identifier.

Complementary Group
(Adjustment Factor Link)

Complementary Group Factor
(Adjustment Factor)

Rather than allowing any possible value as an adjustment factor, 1000 values in 0.5% increments were
created as allowed adjustment factor values through the Complementary Group Factor. The
Complementary Group serves as a three digit representation of the Complementary Group Factor, since
there is an unchanging one-to-one relationship between the two.

 Exhibit 1

Figure 5 - Actuarially Sound Range of CGR Factors and Selection
Two Actuarially Sound Estimates From Class Based Analysis
Premium_1
From Existing Rating Plan

[

]

Premium_Actuarially_Sound = Premium_1 , Premium_2
The actuarially sound range is defined as the values between two
actuarially sound estimates

Premium_2
From New Rating Plan (or Loss Model)

Premiums can also be expressed as Complementary Group Factors (CGR Factors) that result in a given premium
CGR_Factor_1 =

Premium_1
-

Premium_1
CGR Factor that when applied to the
existing rating plan results in
Premium_1

Width of
CGR_Factor_Actuarially_Sound

[

CGR_Factor_Actuarially_Sound = CGR_Factor_1 , CGR_Factor_2

]

CGR Factor that when applied to the existing rating plan results in an
actuarially sound premium between two actuarial sound estimates,
Premium_Actuarially_Sound

Distance
between
CGR_Factor_1

Distance between
Premium_1 and
Premium_2

Selected_CGR_Factor for each micro-segment satisfies all of these conditions
Within the the actuarially sound range as defined above for each micro-segment
Resulting impact (% change in premium) is within a defined range for each micro-segment
Overall rate level on the whole book is maintained (no increase or decrease unless filed)
Total retention on the whole book (as calculated from class based models) is maximized
*Calculations conceptually accurate, but actual calculations take into account the coverage premiums that the CGR factor does/doesn't apply to.

CGR_Factor_2 =

Premium_2
-

Premium_1
CGR Factor that when applied to the existing
rating plan results in Premium_2

Caused by the different estimates of loss, given the
complete set of loss characteristics from two class
based analysis

 Exhibit 2

Background: Factor Selection Process (before CGR)
Assume that a rating plan has 3 variables, each with 3 groupings:
• Territory
• Claims
• Vehicle Type
Proposed factors are between
current and indicated but
movement towards indicated is
based on judgment or an attempt
to mitigate customer impact.
Business considerations have been
and are an acceptable input to the
pricing process.

CONFIDENTIAL AND PROPRIETARY TRADE SECRET - EXEMPT FROM PUBLIC

© Allstate Insurance Company

INSPECTION AND DISCLOSURE

1

 Exhibit 2

Background: Factor Selection Process (before CGR)
If we assume there is one customer
for each unique combination of
rating variables, we can calculate
the current, indicated, and
proposed premiums using the
information on the previous slide.

Example: Territory 1, 1 Claim, Truck
Current Indicated Proposed
Base Rate
$231.48 $223.42 $224.02
Territory X
1.00
0.95
0.99
1 Claim X
1.10
1.10
1.10
Truck
X
1.10
1.30
1.12
Premium = $280.09 $303.51 $273.23

CONFIDENTIAL AND PROPRIETARY TRADE SECRET - EXEMPT FROM PUBLIC

© Allstate Insurance Company

INSPECTION AND DISCLOSURE

2

 Exhibit 2

Background: Factor Selection Process (before CGR)

Using the same
example as before
(Territory 1, 1 Claim,
Truck), we can see
that while this policy
is indicated to
increase by 8.4%
largely due to the
indicated factor for
Vehicle Type, the
proposed factors result
in a negative overall
change.

CONFIDENTIAL AND PROPRIETARY TRADE SECRET - EXEMPT FROM PUBLIC

© Allstate Insurance Company

INSPECTION AND DISCLOSURE

3