The actuary’s nearly three-decade-long status as the sole data experts in their insurance organizations is rapidly changing. Actuaries interact with data scientists more than ever in today’s environment of exponentially growing data and immense computing power, with the lines between the two professions increasingly blurring. The two professions may interact in areas as diverse as claims forecasting models, data architecture and engineering, fraud detection, marketing and program savings evaluation. The points of contact continue to grow, and even for large organizations with distinct departments, the number of dotted lines continues to proliferate.
But just like siblings forced to share a bedroom, the relationship between actuaries and data scientists has not always been smooth. Most poignantly, it has been exacerbated by a lack of recognition of the substantial overlap between the two skill sets coupled with an inability to capitalize on opportunities for collaboration.
To work together productively, actuaries and data scientists must go beyond learning to tolerate each other to learning from each other. Any past rumblings about one profession subsuming the other are more than outdated. Actuaries and data scientists will exist as distinct professions with distinct skill sets while increasingly being asked to work together. Both sides of the professional relationship should engage with this opportunity to reimagine collaboration, leverage each other’s most valuable skills and engage in an open dialogue about how to interact better.
Overlap and Opportunities for Collaboration
Figure 1 outlines some common attributes of the skill sets of data scientists and actuaries. It goes without saying that individuals cannot be fully described by bullet points and boxes, but some broad strokes can be helpful in painting an initial picture. Understanding the comparative strengths, weaknesses and focuses of each profession is a key first step in approaching a more collaborative relationship.
Figure 1: Key Attributes of Data Scientists and Actuaries
There is substantial overlap in the core skill sets of the two professions—ability to apply mathematical/statistical techniques to real-world problems; ability to understand, interpret and explain complex data and models; and facility with technical tools.
Even in areas of distinction, the lines are far from bold. For example, deep industry knowledge and eagerness to adopt change are beneficial within both skill sets while more commonly present in the actuary’s and data scientist’s toolboxes, respectively.
What We Can Learn From Each Other
Actuaries and data scientists have much to learn from each other and should take steps toward capitalizing on each other’s strengths to produce better analysis. This is not a process of blending, but of mutual enrichment. Actuaries do not need to become data scientists, and data scientists do not need to become actuaries (we ought not wish the exam process on even our less-treasured colleagues).
The most important things actuaries can learn from data scientists are:
- Feel free to borrow from other disciplines and try new approaches. Data science is a new field that is not afraid to keep trying new things. A few of the latest developments include genetic algorithms, deep learning and Bayesian adaptive regression trees (BART). Every day that goes by brings a new approach to modeling complex structures. Actuaries should follow and adopt these new tools and approaches when helpful.
- Nonlinear modeling can produce better predictions. Actuaries are accustomed to using factor models or linear regression, and the linear structure is advantageous in terms of interpreting the magnitude of effects (i.e., the expected change in cost for each additional year of age). But when phenomena are truly nonlinear, as is generally the case, linear models are simplistic and not flexible enough to fit well.
- Model validation techniques natural to data scientists are not yet commonly practiced among actuaries, such as receiver operating characteristic (ROC) curves, using train/test/validation samples and variable importance measures, such as the Shapley statistic. These ought to be better understood and used among actuaries, especially because metrics such as area under the ROC curve (AUC)/ROC curves provide new frameworks for examining model accuracy.
- Don’t be afraid of enormous data sets and new variables. Unlimited computing power and ubiquitous data collection mechanisms (e.g., InsurTech) can give insurers a much richer characterization of their members than they could ever get from just looking at demographics and claims.
Likewise, the actuarial profession has proven its value over the test of time, and data scientists could profit from an actuarial perspective in their work.
- The most predictive model is not necessarily the best. An effective model must be communicated to management, the customer or underwriting, and even all three of those parties simultaneously. Business considerations may mean that a simpler model solves the real-life problem better than the most predictive model.
- Don’t start analyzing data until you know it is correct data. Reconciliation of data and creating a tight data set that truly represents your business is a key first step in every project and allows you to draw conclusions later on with greater confidence. In the same way, ensure you fully understand the meaning of the data you are using—data sets are often the instantiation of years of business knowledge and technology debt, so adopt an appropriately humble mindset when diving into that.
- Models need to make sense, at least at some level, to be useful. It’s really not enough to just trust the machine. In fact, basic tests of reasonability are often the keys to identifying underlying errors in modeling. If the model results do not seem reasonable, they are often wrong. An overreliance on modeling techniques can obscure that simple reality unnecessarily.
Actuaries and data scientists can benefit from working together. Though, increasingly, the reality is that they are forced to do so, and turf wars (struggles for responsibility over tasks, responsibilities and positions of influence) are an unfortunately common reality in the early stages of these partnerships. In addition to the common friction that organizational changes can cause, the broad similarities between the two professions can create ambiguity and extra sensitivity when deciding who does what.
RACI, an acronym for “responsible, accountable, consulted, informed,” is a common project management tool that helps align levels of engagement across diverse project stakeholders and can provide a helpful framework for aligning responsibilities and developing an understanding of the unique aspects of each skill set. While at times it is appropriate to formally complete a RACI document, simply being aware of the framework can help guide informal discussions.
For example, a life insurer building a generalized linear model (GLM) to develop premium rates could assign responsibility for building the model to a data science team while asking that actuaries be consulted during that stage of the project to ensure the model will satisfy regulatory scrutiny. Actuaries would then be responsible for the regulatory filings needed to implement the new pricing structure while data scientists would be consulted on model specifics and informed of the final implementation of their modeling work.
One profession does not need to be responsible for all four roles for any one task of a project. In fact, the relative expertise of each profession likely means that it is optimal that both would participate in each task in different roles. At the same time, defining the level of engagement needed from each group at each project stage can prevent the organizational strain of everyone working on everything together.
Working Together—Practical Guidance
While the actuarial and data science skill sets overlap significantly, things as simple as vocabulary can cause unnecessary division. For example, the Society of Actuaries’ (SOA’s) STAM exam tests candidates’ knowledge of the Schwartz Information Criterion, which is universally called BIC (for Bayesian Information Criterion) by data scientists; actuaries use sensitivity analysis to build informal confidence intervals around their estimates, while data scientists use bootstrap sampling and cross-validation; actuaries display data in Excel tables, and data scientists rely on visualizations from R, Python or Tableau.
These are “po-tay-to”/“po-tah-to” differences at the end of the day—that is to say, they are insignificant. It is not hard for either profession to flex their vocabulary toward the other profession—the more significant struggle of being familiar with the concepts the vocabulary describes is already accomplished. Yet, finding a common vocabulary is half the challenge of working together. Similarly, it is incumbent upon a technical professional (actuary or data scientist) to learn to communicate with nontechnical stakeholders, and each should work to remove any vocabulary-based barriers to communication—even among fellow data experts.
Each profession also should seek to understand each other’s work more deeply than at the cursory level. Actuaries should recall how refreshing it is to work with a finance partner who understands the complexities of a reserving process, or a sales partner who seeks to better understand the components of a pricing strategy (beyond “how can we get it lower?”).
For example, actuaries can dabble in using the tools used by their data science partners—they don’t need to be an expert, but something as simple as a weekend spent on Kaggle can be a great immersion experience that yields lasting benefits. Dig deeper than the headlines that read “AI Will Replace All Human Cognition By 2025” (OK, that’s just a touch of an exaggeration). Build your own AI model, play with the parameters and study where it is strong and weak (hint: there are multiple Kaggle tutorials that will lead you through just that).
In the same vein, each profession can empathize with the more difficult aspects of their peers’ daily jobs—this type of support can be uniquely helpful in forming strong professional bonds. Both actuaries and data scientists know the feeling of being asked for a complex data analysis on a tight turnaround only to see it go unused. Both professions know that the 2012 Harvard Business Review article1 proclaiming data science as the “Sexiest Job of the 21st Century” never spent an afternoon digging through code to find the one line in thousands of lines of code that was causing an error. And both can struggle to gain traction communicating complex ideas to nontechnical stakeholders. Work together to overcome those struggles. Share any insights on effective ways to communicate to key stakeholders, partner to build more effective databases, and improve data tools and promote each other’s work wherever possible.
Imagine counseling a college freshman wondering whether data science or the actuarial profession has a brighter future—for many, this situation has been a reality at some point. It is not an either/or question: Data science is vastly more powerful when grounded in deep subject-matter expertise, and actuarial practice is vastly more powerful when enabled by the insights generated by modern methods of data analysis. The future will be marked by the two cooperating in new and innovative ways—far from risking outshining the other, they will grow in tandem as data and analytics remain key components in organizational strategies.
As today’s actuaries chart the future of the profession, we should be mindful of building the future on a foundation of collaboration and mutual understanding while emphasizing a certain actuarial wisdom that has always distinguished an actuarial professional that intuits what benefits are material and cost-effective. Similarly, as today’s data scientists collaborate more closely with actuaries, they can take inspiration from our strong tradition of codified standards of professionalism as well as our strength in analytical reasoning, which is in lock-step with the questions and limitations of the businesses it supports.
Neither profession will subsume the other in the context of insurance organizations. In fact, at many insurers, the two departments are merging, often with data science morphing into a specialized subdepartment of the actuarial group where data science expertise can be harnessed to improve actuarial models and assist an actuarial department in solving a wider variety of insurance problems with substantially more accurate and predictive tools than those in their toolboxes today.
Data scientists will continue to find ample opportunities in a wide variety of noninsurance fields where their techniques add tremendous value. Perhaps expanded collaboration will enable actuaries to lend their analytical and business expertise to the data revolution in noninsurance fields as well.
Copyright © 2021 by the Society of Actuaries, Schaumburg, Illinois.