Predictive Analytics in Business Oliver Werneyer Data and Distribution Leader 19 March 2014 Polish Business Analytics Summit, Warsaw
Everyone's talking about it November 2010 The Economist June 2012 August 2011 2
Less talk. More action Leading work in this area on mortality risk (first morbidity model being built this year) Proven track record in developing propositions for bancassurance Models built for 3 of the top 5 UK bancassurers & launched a predictive product in since 2011 (currently working with the other 2 on building models, launch exp. 2014) Outside the UK, launched a product with HK bank in 2013. Models built for two South African bancassurers launch in next 3 months. Building model for bancassurer in Thailand launch 2014 In addition, a series of client projects where the "intelligent use of data" is a cornerstone of the proposition (health insurance provider, GI and L&H model built, models built on national dataset)
Data analytics in Life Insurance what are the options? Predictive Analytics Two key questions: 1) what do I want to predict? 2) what data do I have access to? Predicting Purchase Predicting Health Predicting lapse Past purchase data available? No past sales data? Do you want to reduce underwriting for the best prospects? Do you want to charge different prices? Do you want to differentiate medical requirements? Past data available on lapsed/nonlapsed customers? No past data available? Build a propensity-to-buy model, which will identify the best prospects for marketing efforts Trigger events can be used (e.g. house move, birth of a child, birthday) Full Predictive Underwriting requires past match-able underwriting & descriptive data (e.g. bank) Model built on mortality/underwr iting data/experience - customers placed in different risk bucket Model selects customers at lowest/highest risk of needing medical tests (e.g. fluids, cotinine) Propensity-tolapse model is built, in order for best products to be sold to, or to direct retention efforts General learnings (e.g. Swiss Re lapse experience) used as starting point (e.g. age/smoker differeces) 4
What do we mean by Predictive Underwriting? The intelligent use of non-medical data held on consumers to reach a view as to their health status These insights can be used to reduce the amount of traditional underwriting (where there is an existing data-rich relationship in place) "You haven't applied for protection, but based on what we know about you, we will pre-approve you and make you an offer" Alternatively, predictive techniques enable you to triage the underwriting process and avoid expensive medical tests for healthy people "Now you are applying for protection, let's run some data on you to remove certain tests, and speed up the process" 5
What does it mean for data to be predictive? Examples of variables found to be predictive of health 6
Building a predictive model Bottom line: any information held on a customer could be predictive of their health status let the data do the talking Crucially, these variables vary by company/market/distribution method, which means that they cannot simply be applied from one company to another Combining all the predictive variables, an algorithm is built that ranks each customer from worst to best prospect, in terms of "likelihood of being given standard rates at application stage" Probability of being a bad risk = 1/(1+e -y ) y = a+bx 1 +cx 2 dx 3 +ex 4 +fx 5 +gx 6 +hx 7 ix 8 +jx 9 kx 10 lx 11 +..+ where: x 1 is age related x 2 is related to value of home x 3 is a brand identifier x 4, x 5, x 7 are related to occupation x 6, x 9, x 11 are account activity related x 8 x 10 are neighbourhood / community related 7
What is needed? Two matchable depersonalised data sources Risk data: c. 50,000 final underwriting decisions from a Life Office The more cases the better Descriptive Data: bank checking account, loyalty card, potentially home/motor insurance The richer the data the better Correlations are found in the descriptive data (the "predictors") Model can be run on whole customer universe to highlight the best prospects 8
Examples of predictive variables Swiss Re has found in bancassurance projects Province / Region Customer Segment Credit Score Other banking products held ATM transactions Income Occupation Property Value Council Tax Bands Household interest (jogging, pets, etc) Age Monthly Income to Outgoings Financial ACORN Group Amount spent on different categories (health products, travel, other insurances) etc. 9
Correlations: income and health 35 Affluence Ranking - Acxiom ILU Characteristics Analysis 18 16 30 25 14 12 10 Likelihood of being rated or declined % (Left Axis) 20 8 6 % of total population (Right axis) 15 4 2 10 Low 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 High 0 10
11 Number of [department store A] Transactions 20% 18% Bad Rate Bad Rate Population Population 70% 60% 16% 14% 50% 12% 40% 10% 8% 30% 6% 20% 4% 2% 10% 0% 0%
12 Sport or Gym Transaction 20% 18% Bad Rate Bad Rate Population Population 80% 70% 16% 60% 14% 12% 50% 10% 40% 8% 30% 6% 20% 4% 2% 10% 0% 0%
Digging deeper Value of Health Transactions in last 12 months (UK Bancassurer) 20 100 18 90 16 80 14 12 70 60 Likelihood of being rated or declined (left axis) 10 8 50 40 Percentage of the population (right axis) 6 4 30 20 Average % of rated or declined across whole population 2 10 0 Females less than X Females over X Males less than X Males over X 0 13
Further refining 18 ATM cash withdrawals in last month (UK Bancassurer) 100 16 90 14 12 10 8 6 4 2 80 70 60 50 40 30 20 10 Likelihood of being rated or declined (left axis) Percentage of the population (right axis) Average % of rated or declined across whole population 0 0 14
Model Output example Fictional model output Cut-off could be set anywhere within this range 100 95 Percentage of standard & substandard 90 85 80 75 Declined 100+ 51-99 Up to 50 Standard 70 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 All Model Output by 5 percentile This tells us, for example, that the top 5% of the model contains a "rated or decline" rate of 5%, as opposed to 14% were no model built (see "All" column) 15
What does the offer look like? Only one health statement to confirm Same price as fully underwritten Max sum assured & term
How does the Predictive Underwriting proposition compare (in the UK market)? Categories Typical paper application fully underwritten Swiss Re Predictive Underwriting proposition Number of questions 20+ 1 5 tick-box Time taken to complete up to 45 mins <5 min Point of sale acceptance up to 40% 100% of those who confirm good health Guaranteed claims payments up to 70% Up to 100% Number of products offered Many 1 Advised/Non advised sale Advised Non advised Issue of non-disclosure High Significantly reduced 17
What are we learning? Key win is for the sales agent: easier process, vastly reduced underwriting, and length of a phone-call consequently dropped from 50m to 7m excellent feedback from telephone agents Post-sale research shows that the biggest influence on decision to buy was #1: Very simple and straightforward policy and documentation #2: It's a pre-approved policy #3: Competitive Premium #4: No need to have a medical exam Our experience is that an element of face-to-face interaction is needed to achieve good response rates A key barrier for banks is getting the "pre-approved process" to the front lines (e.g. in branches) 18
Deployment Direct Marketing Mailings Key Success Factors: Personalised, relevant messaging. Best response rates achievable with multiple response routes, including any of: posted reply, inbound telephony, outbound telephony follow up, referral online, referral into branch. Conventional DM or email. Bank's Role: Provision of data files, any further propensity modelling Expected Response Rates: 0.2-1.0% Outbound Telephony Key Success Factors: Personalised script and targeted to life events Bank's role: Provision of lists, training and call scripting, sales process & monitoring. Expected Response Rates: 1.0-5.0% Online Microsite Key Success Factors: Offer made in secure environment. Tailored offer, optimised by online banking links. Bank's role: Online prompts with link through with microsite developed by 3 rd party. Expected Response Rates: 0.1-0.3% Inbound Telephony Key Success Factors: Cross-sell from other products (credit card, loan etc.) Agent s buy-in and confidence to sell the product. Bank's role: Telephony prompts light touch training and scripting, sales process/monitoring. Expected Response Rates: 1.0-5.0% Easy-to-buy Life Insurance Possible deployments Protection / Life Insurance Consultants Key Success Factors: A time efficient solution that fits well with a normal sales process. Bank's role: Prompts, training/processes and monitoring to align to other protection propositions. Estimated Response Rates: Up to 30% Mortgage Advisers Key Success Factors: Integrated process. Bank's role: Development of Prompts, training/processes and monitoring. Estimated Response Rates: Up to 50% Counter Staff in Branch Key Success Factors: Moving customer away from the public environment (referral to Customer Adviser, online or telephony). Has potential for at counter fulfilment over time, Bank's role: Development of, hand-off processes. Estimated Response Rates: circa 10% of customers who qualify Customer Advisers in Branch Key Success Factors: Buy-in from Customer Advisers when pre-approved customers transact in branch. Bank's role: Branch prompts, light touch training/processes and monitoring. Expected Response Rates: 10-20% ATM Messages Key Success Factors: Call to action to speak to an advisor in branch. Bank's role: ATM Alerts/messages would need to be developed to accommodate this. Estimated Response Rates: unknown 19
How this applies to you 20
For discussion: what's needed for Predictive? Is underwriting currently a barrier to sales (i.e. people put off by the lengthy process/number of questions)? Are you operating in a life insurance market with sufficiently rigorous underwriting? e.g. low levels of non-disclosure as well as a proportion of rated & declined cases Do you and your partners hold data of sufficient quantity and quality for a predictive model build? Do the data laws in your market allow for free transfer of depersonalised data between organisations that are either co-owned (e.g. a bancassurer) or in a strategic partnership (e.g. a joint venture between life office and bank/retail group)? Are both parties (life office and the provider of "other" descriptive data) willing to share their depersonalised data with each other for the benefit of increased sales through Predictive Underwriting? 21
22
Legal notice 2014 Swiss Re. All rights reserved. You are not permitted to create any modifications or derivative works of this presentation or to use it for commercial or other public purposes without the prior written permission of Swiss Re. The information and opinions contained in the presentation are provided as at the date of the presentation and are subject to change without notice. Although the information used was taken from reliable sources, Swiss Re does not accept any responsibility for the accuracy or comprehensiveness of the details given. All liability for the accuracy and completeness thereof or for any damage or loss resulting from the use of the information contained in this presentation is expressly excluded. Under no circumstances shall Swiss Re or its Group companies be liable for any financial or consequential loss relating to this presentation. 23