This Special Issue of the North American Actuarial Journal (NAAJ) contains nine articles that contribute to the advances in predictive analytics with a wide range of applications in insurance and risk management. Draft versions of the articles were presented at the Advances in Predictive Analytics (APA) conference, which was held on December 1–2, 2017, as part of Faculty of Mathematics’ 50th anniversary celebration at the University of Waterloo, Waterloo, Ontario, Canada. The conference was co-hosted by the Department of Statistics and Actuarial Science and the Waterloo Research Institute in Insurance, Securities and Quantitative Finance of the University of Waterloo and co-sponsored by the Society of Actuaries’ Center of Actuarial Excellence and the Department of Statistics and Actuarial Science, University of Waterloo. The conference was able to bring leading academic researchers and practitioners together in one place to share recent advances in predictive analytics related to a wide range of applications in insurance and risk management, as well as practical experiences and best practices on a variety of topics around this theme in the insurance industry.
The conference was addressed by the following four keynote speakers:
Patrick Brockett, University of Texas at Austin, gave a presentation titled “Rapid Estimation of Disaster Relief Fund Distribution: Iterative Learning with Diverse Geospatial Data Inputs.”
Ian Duncan, University of California Santa Barbara, gave a presentation entitle “Healthcare: How Can We Harness Predictive Analytics for Patients, Providers and Payers?”
Edward W. (Jed) Frees, University of Wisconsin-Madison, gave a presentation titled “Predictive Analytics and Medical Errors.”
Greg Taylor, from the University of New South Wales, gave a presentation titled “The Long Road to Enlightenment: Loss Reserving Models from the Past, with Some Speculation on the Future.”
In addition to the above invited keynote academic presentations and 20 (academic) contributed presentations, the conference was highlighted with two practitioners’ panel discussion sessions, one case study session, as well as speeches by Sharon Giffen (past president of the Canadian Institute of Actuaries) and Jerry Brown (past president of the Society of Actuaries). The two practitioners’ panel sessions were organized by Ben Marshall (Society of Actuaries) and Chris Fievoli (Canadian Institute of Actuaries). Below are some details of the two practitioners’ panel sessions:
Panel Discussion in Life Insurance
Session description: This panel offered practitioner perspectives on applications of predictive analytics in the life insurance industry. Speakers addressed questions such as: What are the key success factors for a life insurance predictive analytics team and a data scientist? Is an actuarial data scientist different than other data scientists? What are the typical hurdles in setting up an insurance analytics function? How do you maintain the balance between organizational agility and compliance? What types of techniques are used for solving life insurance problems with analytics? How much effort is typically spent on preparing/cleaning the data? What are the key metrics used to assess the predictive power of a model?
Panelists: Ian Bancroft (Sun Life Financial), Kevin Pledge (Claim Analytics), Jean-Yves Rioux (Deloitte Canada), and Eugene Wen (Manulife Financial).
Panel Discussion in Property and Casualty Insurance
Session description: Predictive analytics is gaining increased importance in the property and casualty insurance industry, as property and casualty companies are looking for ways to better integrate diverse sources of data into their operations. Our panelists, each a practitioner in this area, presented their insights on how predictive analytics were being used in their own companies and how future research initiatives could potentially help.
Panelists: Jeffrey Baer (Economical Insurance), Denise Cheung (Aviva Canada), Sébastien Bernard (Intact Insurance), and Frédérick Guillot (The Co-operators Insurance and Financial Services).
The case Study Session was organized by Chris Fievoli from the Canadian Institute of Actuaries. The speakers of the session were Jeffrey Baer (Economical Insurance), Dragos Capan (Manulife Financial), and Denise Cheung (Aviva Canada).
As an integral part of this conference, presenters were encouraged to submit their full papers to be considered for publication in NAAJ’s Special Issue, subject to the regular referee process. Below we briefly describe the nine papers that are appearing in this special issue.
In “Data Clustering with Actuarial Applications,” Guojan Gan and Emiliano A. Valdez first provide a review of the basics of data clustering, such as distance measures and cluster validity, and different types of clustering algorithms. Then they demonstrate the applications of data clustering in insurance by using two scalable clustering algorithms, the TFCM algorithm and the hierarchical k-means algorithm, to select representative variable annuity contracts that were used to build predictive models. They find that the hierarchical k-means algorithm is efficient and produces high-quality representative variable annuity contracts.
The article “Efficient Nested Simulation for Conditional Tail Expectation of Variable Annuities,” by Ou Dang, Mingbin Feng, and Mary R. Hardy, a simulation procedure for estimating the conditional tail expectation (CTE) of liabilities of a variable annuity dynamic hedging strategy is proposed. In a CTE calculation, tail scenarios—that is, the scenarios that result in extreme losses—are most relevant. In correctly identifying those scenarios, they show that this would improve the efficiency in a nested simulation. The propose a procedure that takes advantage of the special structure of the CTE by first identifying a small set of potential tail scenarios from the first tier of simulation. Then they focus the simulation budget on only those scenarios. They carry out an extensive set of numerical experiments on different guarantee types and different stochastic stock return dynamics. Their numerical results show that, when given a fixed simulation budget, the proposed procedure could improve the accuracy of CTE estimation by an order of magnitude compared to a standard nested simulation.
The article “Predictive Analytics and Medical Malpractice” by Edward W. Frees and Lisa Gao is a case study on the use of predictive analytics in the context of medical errors. The authors analyze medical errors using data external to some health care systems for an improvement of the systems through medical malpractice insurance. In the spirit of modern analytics, they describe the application of data from several different sources that provide different insights into a specific problem facing the medical malpractice community: the relative importance of upper limits (or caps) on insurance payouts for noneconomic damage (e.g., pain and suffering). This topic is important to the industry in that many courts are considering the legality of such limitations. All stakeholders, including patients, physicians, hospitals, lawyers, and the general public, are interested in the implications of removing limitations on caps. The authors demonstrate how one can use data and analytics to inform the many different stakeholders on this issue.
In the article “Drivers of Mortality Dynamics: Identifying Age/Period/Cohort Components of Historical U.S. Mortality Improvements,” authors Johnny S.-H. Li, Rui Zhou, Yanxin Liu, George Graziani, R. Dale Hall, Jennifer Haid, Andrew Peterson, and Laurence Pinzur set out to obtain an age/period/cohort (A/P/C) decomposition of historical U.S. mortality improvement. Two different routes to achieving this goal are studied. In the first route, the desired components are obtained by fitting an A/P/C model directly to historical mortality improvement rates. In the second route, an A/P/C model is estimated to historical crude death rates and the desired components are then obtained by differencing the estimated model parameters. For each route, various possible A/P/C model structures are experimentally investigated and evaluated on the basis of their robustness to several factors (e.g., changes in the calibration window) and their ability to explain historical changes in mortality improvement. Based on the evaluation results, an A/P/C decomposition for each gender was recommended.
In “Pricing Flood Insurance with a Hierarchical Physics-Based Model,” by Mathieu Boudreault, Patrick Grenier, Mathieu Pigeon, Jean-Mathieu Potvin, and Richard Turcotte, a hierarchical approach to generate a set of simulations of flood losses from the Chaudiére River (Québec) is proposed. Firstly, they use scenarios from a chain of physically based models representing climate, hydrology, and hydraulicity as well as simulations from an econometric model to represent water discharge at a specific location on the river. Secondly, they utilize the distribution of water discharge in a hydrologic/hydraulic model (HEC-RAS) to calculate water levels (stage) in a 3D map along the river. In addition, they employ civil engineering data to map the level of damage as a function of water level. Finally, they determine the distribution of the total loss for a portfolio of those insured as well as the distribution of the loss for each individual policy via Monte Carlo simulation.
In the article “Efficient Simulation Designs for Valuation of Large Variable Annuity Portfolios,” authors Ben Mingbin Feng, Zhenni Tan, and Jiayi Zheng provide a set of comparisons among recently proposed machine learning methods for large variable annuity portfolio valuation. Specifically, they discuss pitfalls of some of the methods and propose corresponding improvements. In addition, they propose, analyze, and test a new valuation method based on their suggested improvements. They show that the resulting procedure achieved both higher accuracy and lower computational requirements than the previously proposed methods.
In the article “Bühlmann Credibility-Based Approaches to Modeling Mortality Rates for Multiple Populations,” by Cary Chi-Liang Tsai and Adelaide Di Wu, three non-parametric Bühlmann credibility-based approaches to modeling multi population mortality rates are proposed, including the joint-k Bühlmann credibility model, the co-integrated Bühlmann credibility model, and the augmented common factor Bühlmann credibility model. They use mortality data for both genders in the United Kingdom, the United States, and Japan from the Human Mortality Database. Mortality data for a selected age span and each of a series of year spans for two populations (both genders in each country) and six populations (both genders in all three countries) are first fitted to each of the six models (three Lee-Carter-based and three Bühlmann credibility-based). Next, they compare the forecasting performances of the six underlying mortality models with the measure of mean absolute percentage error between the predicted and observed mortality rates for three forecasting time spans (10, 20, and 30 years).
The article “Predictive Modeling of Threshold Life Tables,” by Min Ji, Mostafa Aminzadeh, and Min Deng, considers several parametric prior distributions as well as Jeffreys priors to derive an appropriate predictive density of lifetime random variable. They use the predictive density to compute life expectancy and other measures of interest in a Bayesian framework. They also provide simulation studies to assess the accuracy of the estimates of interest based on the predictive density.
The final article, “Remote Sensing Applications for Insurance: A Predictive Model for Pasture Yield in the Presence of Systemic Weather,” by C. Brock Porth, Lysa Porth, Wenjun Zhu, Milton Boyd, Ken Seng Tan, and Kai Liu, provides a comprehensive comparison of 13 pasture production indices, including those developed based on satellite-derived vegetation and biophysical parameter indices with regard to their effectiveness in modeling pasture yield and index-based insurance policy. This research makes an important contribution to the field of actuarial science and insurance, because it highlights potential new opportunities for insurance design and predictive analytics using large and comprehensive satellite datasets that remain relatively unexplored to date in insurance practice. Though pasture insurance was used in their analysis, the research could be extended to other crops and other areas in the property and casualty sector, including fire and flood.