Predictive Analytics for Profitability
HOW TO USE PREDICTIVE ANALYTICS TO TURN DATA INTO ACTION
Is Big Data the Answer?
We spend an awful lot of time talking about data in insurance, especially big data. Big data is the thing. Big data is huge. Big data is the best thing since the last best thing we had.
When something has so much potential and so much attention, it’s easy to get hung up on the hype. But how is more data going to make insurance organizations successful? Is the data itself going to get us where we need to be? Probably not.
The answer to what?
At a recent industry meeting we talked for two days about all the new data now available, what is the right data to collect and use, and how to get the data right. We never got to the agenda items covering the right way to analyze data and apply it to business problems. Or even what those business problems are.
We spend a lot of time talking to insurance executives, and it seems most of their strategic problems boil down to one of these two themes:
• How to maximize profit without sacrificing growth • How to maximize growth without jeopardizing profit
We know the dials to turn to increase growth – sales channels, marketing efforts. We know the dials to turn to increase profit – expenses, retention. What we have trouble with is the balancing act, making sure turning up one doesn’t push the other down. The result can be underpriced policies that kill loss ratios and overpriced policies that walk away.
We don’t believe big data is the answer to everything, but with the right tools and approach, predictive insights gained by insurance data mining can be the answer to the growth/profitability dilemma.
What happens when you act on
predictive analytics insights?
• React more quickly to market changes. • Reduce claims expenses.
• Optimize portfolio performance. • Build a culture of data-based
Making Sense of Rate Inaccuracies
Here’s a good example of the growth/profitability dilemma in real life. Even in the personal auto insurance space – the line of business most advanced at using data and analytics – pricing inaccuracy can be glaringly obvious.
Almost everyone has seen the popular series of video ads with onscreen comparisons showing the advertiser’s rates far lower than the other brands. This YouTube commenter expresses what all consumers are probably thinking when they see these ads.
“Trwent” is probably right. The rates shown in these ads and in online car insurance quoting tools can’t all be accurate for each customer.
When we tested the accuracy of some actual auto quotes with a set of rates from multiple carriers on three policies, the results were stunning.
For Policy 1, rates range from $349 to $802 – a difference of $453. Policy 2 has a similar range. For Policy 3, the range of quotes is staggering – from $764 to $3,582.
When rate disparities are this large, it’s clear that some of the rates aren’t accurate, even though they are set by using essentially the same tools with similar methodologies. Each brand would probably tell you their rates are accurate, maybe off by ±5%. But this is not 5%.
If the winning brand in the Policy 3 quoting war is the lowest price brand, then that brand will most likely be the biggest loser when it comes to the policy’s ultimate loss ratio.
And if auto insurance carriers using not only massive volumes of data but also incorporating leading-edge data like telematics can’t get it right, then how can anyone else? Why isn’t all this new data improving their pricing accuracy? Why can’t they do a better job of predicting the future profitability of new business?
Brand
1
2
3
4
5
6
7
8
9
10
11
12
Policy 1
$596
$373
$463
$777
$604
$381
$398
$802
$349
$479
$676
$555
Policy 2
$0
$292
$292
$599
$560
$253
$229
$360
$306
$306
$326
$416
Policy 3
$2,015
$1,909
$2,278
$2,713
$0
$1,022
$764
$2,135
$3,582
$0
$0
$952
The Right Tool to Do the Job Right
Every home improvement fanatic knows you have to have the right tools to do a job right. The same applies to mining data for predictive insight – start with the right tools.
The insurance industry isn’t exactly notorious for early adoption of new technology. We’ve been very dependent on spreadsheets, powerful as they are in the right hands, for more than two decades. But spreadsheets are insufficient for extracting predictive insights from data.
On the other hand, the pace of adoption for Business Intelligence (BI) and Data Visualization tools has increased dramatically since the data explosion began. But those tools are best at allowing us to slice and dice views of what happened in the past, and are, like spreadsheets, very limited in their predictive power. Popular tools like Generalized Linear Modeling and Min-Bias techniques help determine the right rating factors without double-or triple-charging a single risk, for example a 22-year-old single male driver with bad credit and multiple accidents. The GLM spreads the risk among the different factors. But because the GLM can’t identify interactions in the data, 22-year-old single males with bad credit and multiple accidents get treated as a distinct group. The power of that variable combination is then diluted instead of being spread out. Perhaps it’s time we looked to market leaders in other industries for a new idea.
Machine Learning Drives the Best Predictive Modeling Tools
Machine learning selects your next streaming movie or song. Hedge funds use it for automated trading. It’s the decision engine in driverless vehicles. It’s faster, more accurate, and more predictive than other analysis tools. Machine learning is automated modeling that learns as it goes, applying algorithms in an iterative fashion. It searches through hundreds and thousands of data interactions, creating and testing model after model to ultimately produce the most reliable predictive model possible from the data.
Machine learning isn’t a new concept, but new applications of machine learning are changing the fortunes of nearly every industry. Need proof? Watch Netflix.
Consider how machine learning works in online movie streaming. We don’t know Netflix’s algorithms, but it’s easy to observe what happens. As you watch movies, Netflix tracks what you watch, what you abandon, what you put in your queue and whether you watch it. It uses that information to present viewing suggestions for you – and for other people with characteristics similar to yours.
Viewing suggestions aren’t the only use of machine learning for a company like Netflix. Companies in every industry can use machine learning to identify the attributes of customers who reliably pay their bills (or don’t), and target their marketing accordingly. The possibilities are endless.
With machine learning,
insurers can capitalize on
big data to outcompete in
every area:
• Out-segmenting the competition in pricing and underwriting • Finding the best customers • Managing claims cost-effectively
Machine Learning Applied to Insurance
Machine learning works the same way for insurance. Using data on existing policies and claims, machine learning algorithms find the interactions of variables associated with a stated result, such as retention or loss ratio.
Machine learning is equally effective for large and smaller insurers. While it efficiently handles vast volumes of data, it can also produce highly credible models from smaller data sets
Because machine learning is automated, with massive computing power, it can model huge quantities of data very quickly. In any data set, hundreds and even thousands of possible data interactions exist, and machine learning searches through all of them. It learns from each repetition, or iteration, refining until a final model is produced to represent the most predictive signal in the data.
Could a person try to do this? Yes, and they do, to some extent, with Generalized Linear Modeling. But the approach is largely trial and error, and depends heavily on human intuition pointing the model in the right direction. Machine learning simply uses computers to do what computers are good at doing: Checking out the many possible combinations. Machine learning also lets humans do what they are good at doing: Interpreting what the model is and isn’t telling you and deciding what to do next.
Let’s look at some examples where machine learning found predictive insights that were immediately actionable.
Turning Predictive Insight Into Actionable Insight
Example 1Analysis goal: Who are our most profitable customers? What are the characteristics we should look for in
new business?
Predictive insight: In a data set from a private passenger auto carrier with a total portfolio loss ratio of 71%,
even without machine learning it was easy to see that policies where a safe driver discount was applied had a loss ratio of 87%.
Considering those without the discount had a far lower loss ratio of 59%, it seems clear that the carrier was over-discounting.
Using machine learning to analyze the variable interactions, many more refined segments emerged. One segment stood out for its extremely low loss ratio compared to the 71% portfolio loss ratio.
A segment of policies with the safe driver discount that also had 10 years or more driving experience, but did NOT have a passive restraint discount, actually had a surprisingly low 35% loss ratio. The segment would have been difficult to find or test using traditional methods, and was a credible segment that was
statistically valid.
Action: If the carrier reduced everyone’s safe driver
discount, many drivers who have the discount would be penalized and may shop for a new policy.
Instead, the carrier has implemented a loss ratio scoring process to target this valuable segment, and others with characteristics that are predictive of low loss ratios, to grow new business and improve retention.
Safe driver discount:
Yes
No
Loss ratio:
87%
59%
No
Yes
59%
87%
76%
51%
74%
58%
10+
No
Safe
driver
discount
Loss ratio
=35%
Tenure
Passive
restraint
discount
Yes
<10
Example 2
Analysis goal: Identify the predictors of the worst loss ratios so potential “loss ratio killers” can be
identified and avoided.
Predictive insight: In another example from the same data set, policies with one or more male driver had a
higher-than-average loss ratio of 75%.
Policies with drivers with less than two years’ experience had an even higher loss ratio of 90%.
Neither of these is surprising, and a logical next step would be to look at the different combinations: Less than 2 years with no male drivers vs. <2 years with male drivers, 2+ years with no male drivers vs. 2+ years with male drivers, etc.
Machine learning takes that next step, and many more, by looking at combinations of all variables included in the data set.
Analyzing not only these two variables (tenure and # male drivers) but also adding three more variables yields a segment with a 150% loss ratio:
• One or more male drivers
• Less than two years’ driving experience
• One or more unmarried driver
• Minimum driver age less than 50
• Vehicle age six or more years
Action: Using the model to identify high loss ratio
policies at the point of quote could allow the insured to avoid unprofitable business.
Male drivers:
1+
0
Loss ratio:
75%
66%
Driving tenure (yrs)
<2
2+
Loss ratio:
90%
63%
<2 yrs
2+ yrs
90%
63%
<50
50+
78%
60%
62%
83%
1+
Loss ratio
=150%
Tenure
Unmarried
drivers
Min
driver
age
Male
drivers
Vehicle age
0
66%
75%
1+
0
75%
69%
6+ yrs
<6 yrs
Example 3
Analysis goal: Create a model to identify
large losses as early as possible after a claim is reported.
Predictive insight: Let’s look at one more
example, this time in liability claims. Claims organizations everywhere have their checklists of the flags that most frequently indicate large loss potential. These checklists have developed over many years of experience.
This carrier set a threshold of $100,000 as defining a large loss. Looking back at the data after claims closed, we could see that at just 10 days after the claim was reported the carrier’s method had identified 12% of the large losses. Using machine learning on the same set of claims, however, identified 65% of the large losses using the information available at the 10-day mark.
Action: Every week, the insurer runs all open claims through an early-warning claims severity model. Claims
flagged go to adjusters with the appropriate skill level and authority to apply claims management resources. This approach ultimately saves 5% to 10% on average for large losses – a savings that can add up quickly.
Predicting the Future of Insurance
It doesn’t take machine learning to see that every day it becomes clearer that insurers who do not embrace analytics to mine predictive insight from data – and then do something about it – are going to be left behind by competitors who do. Some are way ahead on this journey, and others haven’t gotten started.
Many are stalled on their data analytics journey by questions about data quality, or distracted by the potential of big data. While those factors are definitely barriers to extracting predictive analytics using traditional tools, with machine learning, you can still get started now, with existing data, and find actionable insight that could make the difference in the growth vs. profitability balancing act.
% of large
losses
70%
60%
50%
40%
30%
20%
10%
0%
Identified by carrier Identified by machine learningCelina Insurance Reduces Expenses and
Improves Processes Across the Business
Personal Lines Use Case
Challenge
Facing profitability issues, regional property and casualty carrier Celina Insurance Group’s Chief Actuary couldn’t help but notice the prevalence of predictive analytics use among his peers, especially at the national carriers with whom Celina competes.
Celina determined that predictive analytics could provide solutions to their short-term expense reduction needs as well as their longer-term profitability challenges. In the process, they could achieve a cultural change to become more analytics-driven across the organization.
Solutions
Celina embarked on a strategic vendor selection process, which is documented in the Celent report, Celina Insurance’s Predictive
Analytics Initiative: The Machine Learning Factor.
Celina initially engaged Guidewire to segment their Homeowners portfolio to identify profitable risk groups, as well as to address any inadequate pricing revealed by the segmentation.
As Celina gained confidence using the system, they completed a Motor Vehicle Report (MVR) expense reduction project intended to help them master predictive modeling techniques and deliver quick ROI on the software investment.
Early results
Segmentation of Personal Auto & Homeowners Business: Celina’s first business applications of the predictive analytics system were to their Homeowners and Personal Auto programs, which make up 63% of premium volume. Using company data, Celina identified segments of the business that were either over- or under- priced by the existing rating structure. Rating factors to address pricing deficiencies were put into effect in each states’ homeowners and standard and preferred personal auto programs.
MVR Expense Reduction: Using Guidewire Predictive Analytics
software, Celina modeled the auto policies where Motor Vehicle Reports made a significant difference in underwriting outcome, identifying segments where the reports didn’t add value. Celina targeted savings of approximately $10,000 per month by limiting MVR orders; their actual savings exceeded $22,000 per month.
Summary
Regional carrier needed to improve its ability to compete with national carriers in all lines.
Early results:
• A homeowners and auto
segmentation project generated new rates that were approved in all four personal lines states. • An MVR expense reduction
project delivered fast ROI and savings of more than double the targeted results.
Additional applications:
• Risk scoring on 2/3 of their book of business provides underwriting quality scores and pricing scores.
• An agency appointment score guide will help more strategically identify areas to appoint new agents.
About Celina Insurance Group Founded in 1914, Celina Insurance Group is composed of four mutual property and casualty insurance companies. Celina underwrites auto, non-standard auto, farm, home, commercial auto and property insurance, as well as umbrella and casualty coverage in six Midwestern states. Celinainsurance.com
Additional Predictive Analytics Applications at Celina
Celina has continued to increase the sophistication of their predictive modeling techniques, and has integrated real-time risk scoring into the underwriting process for two-thirds of the total book of business. With the scoring implementation, new business and renewal underwriters see a frequency score and a pricing score for each risk. These provide invaluable information to measure the quality of business and improve underwriting accuracy. Celina is closely monitoring the results of these changes, and expects to be able to share quantified results in the near future.
After seeing the positive impact that predictive analytics has had in pricing and underwriting, Celina’s Marketing team also decided to integrate predictive analytics into their agency appointment strategy. A new model based on key performance indicators for five programs including auto, homeowners,
and commercial business ranks the 4,200 zip codes in which Celina writes business. By targetting agency appointments in the zip codes with the highest performance indicators, Celina can pursue profitable growth opportunities in previously under-represented areas.
“The Marketing team’s
request to develop a guide to
help find areas for profitable
growth, then marry that data
with agency appointment
strategies, was a clear example
of how the entire organization
fully appreciates the power of
advanced analytics to inform
better decision making.”
— Vince Franz, Chief Actuary, Celina Insurance Group
Buckeye Insurance Group Achieves
Sustainable Profitable Growth
Personal Lines Use Case
Challenge
Buckeye’s auto business was struggling to grow and remain profitable. They knew they needed to take rate action, but wanted to go beyond the traditional rate revision process. Buckeye wanted to take a leap forward with their pricing segmentation and sophistication, not just catch up to the market.
Solution
Buckeye contacted their reinsurance broker, Guy Carpenter, in search of a predictive analytics solution. After evaluating several vendors, they chose Guidewire Predictive Analytics for both their Machine Learning software platform and actuarial consulting expertise.
Buckeye and Guidewire partnered to create an innovative rating program. The new rate structure
used standard rating factors derived with Generalized Linear Modeling (GLM), combined with new rating tables created by Guidewire’s proprietary Machine Learning algorithms.
The GLM laid down an elementary base for Buckeye’s rating structure and the Machine Learning discoveries took them to the next level by incorporating complex variable interactions. These variable interactions described fully credible segments of loss characteristics and became new rating factors, allowing Buckeye to price and select risk more profitably. These risk patterns had always existed in their data, but machine learning technology made their discovery possible.
Results
Guidewire completed the data preparation and modeling in just three months. The Machine Learning analysis uncovered significant segments of underpriced business, and Buckeye chose to take targeted rate action. The new rating plan was launched after receiving Department of Insurance approval in each of their states. One year after implementation, Buckeye’s written premium had increased by more than 10% while their loss ratio decreased by 4 points. Key Performance Indicators (KPIs) pointed to sustainable success:
What’s next?
Buckeye is already expanding its relationship with Guidewire to implement new rating plans for other lines of business. In addition, they are deepening their machine learning sophistication with scoring. The expected result: Buckeye will move even further ahead of their competition in pricing accuracy, with granularity that allows for more accurate insight into individual risks and less rate volatility when policy
Summary
1. Auto profits were declining.
2. Buckeye wanted not just to update rating structure but to take a leap forward
3. Guidewire applied proprietary Machine Learning to the data.
4. Resulting new rate tables were approved in all of Buckeye’s states.
5. One year after rate changes, written premium increased by more than 10% while loss ratio decreased by 4 points.
6. Buckeye returned to profitable growth and KPIs point to sustainable profitability.
About Buckeye
Ohio-based Buckeye Insurance Group specializes in providing personalized insurance coverage to customers living in rural communities in several states.
• Top-line premium increased.
• New business conversion improved.
• Loss ratio decreased.
• Retention remained steady to slightly up.
-5 pts loss ratio
+10% written premium
Large Auto Carrier Adds 27% Lift to GLM Model
Personal Lines Use Case
Challenge
The insurer, which is the largest domestic property and casualty insurance company in its state, decided to completely revise their rating plan and introduce it as a new program. In designing the new rating plan, all factors were considered:
• New base rates
• Rating factors
• Territory definitions and relativities
Solution
The approach incorporated the strengths of both Guidewire GLM and Guidewire Complete Segmentation.
Guidewire GLM was used to build an initial model for the bulk of the class plan and to control for certain other fields (e.g. geographic effects). Residuals from this initial GLM were analyzed by Guidewire complete segmentation to discover the compound variables undetected by the GLM framework.
Given the new compound variables along with the initial GLM, Guidewire complete segmentation was used again to explore geographic fields to determine new territory definitions.
Results
This project was extensive, involving territory definitions, base rates, and rating plan factors for five different coverages. The timeline spanned approximately six months using the powerful Guidewire Predictive Analytics Suite.
The complete segmentation analysis added 27%
lift to the GLM model, with an average correlation of 95.8%, indicating a very high degree of consistency on unseen validation data.
The new rating plan was filed and approved by the state’s Department of Insurance. With more accurate rates in place, the insurer is positioned for profitable growth.
Guidewire Advantages
Guidewire Predictive Analytics’ GLM uniqueness and superior results derive from its integration with our powerful complete segmentation capabilities. The combined solution streamlines and accelerates the process of building rating plans while providing greater accuracy than traditional GLMs.
+27% lift over GLM 100 80 60 40 20 0 Cumulative estimates (%)
Lorenz curve from initial GLM
20 40 60 80 100 120 50 40 30 20 10 0 Loss ratio %
Segmentation model for territory definitions
Segment
Auto Carrier Uses Segmentation to
Reduce the New Business Penalty
Personal Lines Use Case
Challenge
A large Canadian Personal Auto insurer wanted to improve profitability on new business writings.
Solution
Guidewire suggested using a new business quote filter that would gradually eliminate the most unprofitable risks from the insurer’s portfolio. Using a complete segmentation on policy level data with the loss ratio as the objective, the expected profitability of any risk can be determined accurately. The insurer can achieve the desired profitability increase by filtering out the highest expected loss ratio segment.
Results
The data preparation and modeling exercise were completed within 90 days of data receipt. The quote filter was implemented by the insurer within two months of the final presentation, with the results then being monitored closely to ensure the filter was working properly.
Four months after implementation, the proportion of highest loss ratio risks in the new business portfolio decreased from 8 points to 3 points, which translates into a 5 point improvement in expected loss ratio for new business.
Guidewire Advantages
By using machine learning methods, the model development process is much faster compared to traditional approaches, such as GLMs. Multiple iterations, using different allowable predictors and model complexities, can be built to account for implementation or regulatory constraints. The segments derived by Guidewire Predictive Analytics have clear, easy-to-understand descriptions that guide acceptability of the model.
1.04 1.02 1.00 0.98 0.96 0.94 0.92 0.90
Expected loss ratio
Month
Quote filter
implemented
Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov Jan Mar May Jul Sep Nov
-5 pts
loss ratio
© Guidewire Software, Inc. All rights reserved. Guidewire, Guidewire Software, Guidewire PolicyCenter, Guidewire ClaimCenter, Guidewire BillingCenter, and the Guidewire logo are trademarks or registered trademarks of Guidewire Software, Inc. in the United States and/or other countries. GW-PR-PA-20160418
About Guidewire
Guidewire delivers the software that Property/Casualty (P/C) insurers need to adapt and succeed in a time of rapid industry change. We combine three elements – core processing, data and analytics, and digital engagement – into a technology platform that enhances insurers’ ability to engage and empower their customers and employees. More than 200 P/C insurers around the world have selected Guidewire. For more information, please visit www.guidewire.com and follow us on twitter: @Guidewire_PandC.
