April 24, 2014
The Science (and Art) of Data, Part 2
by Anand Rao
There are not enough good data scientists to go around. So, should you "buy" them, "rent" them or "build" them. A hybrid may be the answer.
Given the high need and growing demand for data scientists, there are definitely not enough of them. Accordingly, it is important to consider how an insurer might develop a core talent pool of data scientists. As it is often the case when talent is in short supply, acquiring (i.e., buying) data scientist talent is an expensive but fairly quick option. It may make sense to consider hiring one or two key individuals who could provide the center of gravity for building out a data science group. A number of universities have started offering specialist undergraduate and graduate curricula that are focused on data science, which should help address growing demand in relatively soon. Another interim alternative is to “rent” data scientists through a variety of different means – crowdsourcing (e.g., Kaggle), hiring freelancers, using new technology vendors and their specialists or consulting groups to solve problems and engaging consulting firms that are creating these groups in-house.
The longer term and more enduring solution to the shortage of data scientists is to “build” them from within the organization, starting with individuals who possess at least some of the necessary competencies and who can be trained in the other areas. For example, a business architect who has a computational background and acts as a liaison between business and technology groups can learn at least some of the analytical and visualization techniques that typify data scientists. Similarly, a business intelligence specialist who has sufficient understanding of the company’s business and data environment can learn the analytical techniques that characterize data scientists. However, considering the extensive mathematical and computational skills necessary for analytics work, it arguably would be easier to train an analytics specialist in a particular business domain than to teach statistics and programming to someone who does not have the necessary foundation in these areas.
Another alternative for creating a data science office is to build a team of individuals who have complementary skills and collectively possess the core competencies. These “insight teams” would address high-value business issues within tight time schedules. They initially would form something like a skunk works and rapidly experiment with new techniques and new applications to create practical insights for the organization. Once the team is fully functional and proving its worth to the rest of the organization, then the organization can attempt to replicate it in different parts of the business.
However, the truth is there is no silver bullet to addressing the current shortage of data scientists. For most insurers, the most effective near-term solution realistically lies in optimizing skills and in team-based approaches to start tackling business challenges.
Designing a data science operating model: Customizing the structure to the organization’s needs
To develop a data science function that operates in close tandem with the business, it is important that its purpose be to help the company achieve specific market goals and objectives. When designing the function, ask yourself these four key strategic questions:
- Value proposition: How does the company define its competitive edge? Local customer insight? Innovative product offerings? Distribution mastery? Speed?
- Firm structure: How diverse are local country/divisional offerings and go-to-market structures, and what shared services are appropriate? Should they be provided centrally or regionally?
- Capabilities, processes and skills: What capabilities, processes and skills do each region require? What are the company’s inherent strengths in these areas? Where does the company want to be best-in-class, and where does it want to be best-in-cost?
- Technology platform: What are the company’s technology assets and constraints?
There are three key considerations when designing an enterprisewide data science structure: (a) degree of control necessary for effectively supporting business strategy; (b) prioritization of costs to align them with strategic imperatives; and (c) degree of information maturity of the various markets or divisions in scope.
Determining trade-offs: Cost, decision control and maturity
Every significant process and decision should be evaluated along four parameters: (a) need for central governance, (b) need for standardization, (c) need for creating a center of excellence and (d) need for adopting local practices. The figure below illustrates how to optimize these parameters in the context of cost management, decision control and information maturity.
This model will encourage the creation of a flexible and responsive hub-and-spoke model that centralizes in the hubs key decision science functions that need greater governance and control, and harnesses unique local market strengths in centers of excellence. The model localizes in regional or country-specific spokes functions or outputs that require local market data inputs, but adheres to central models and structures.
Designing a model in a systematic way that considers these enterprise-wide business goals has several tangible benefits. First, it will help to achieve an enterprisewide strategy in a cost-effective, timely and meaningful way. Second, it will maximize the impact of scarce resources and skill sets. Third, it will encourage a well-governed information environment that is consistent and responsive throughout the enterprise. Fourth, it will promote agile decision-making at the local market level, while providing the strength of heavy-duty analytics from the center. Lastly, it will mitigate the expensive risks of duplication and redundancy, inconsistency and inefficiency that can result from disaggregation, delayed decision making and lack of availability of appropriate skill sets and insights.