Success with Predictive Analytics—Getting the Basics Right

Data has been deemed the oil that will fuel the next industrial revolution. Enterprises have the opportunity to leverage data at their disposal, be it transactional enterprise data or unstructured data in the form of emails, weblogs, Internet of Things (IoT) data streams, or clickstreams. Data represents a way not just to understand current user preferences, industries, and business environments, but also to predict future changes.

Traditionally this would be done by analyzing historical data. For instance, by analyzing online sales a retailer can better target promotions based on which users are more likely to purchase a particular product. Factory managers can forecast when industrial equipment will fail by detecting patterns in machine and sensor logs. This branch of advanced analytics, which uses historic data to predict future events, is referred to as predictive analytics.

Today, with advances in technology predictive analytics leverages structured enterprise data and unstructured data, then uses techniques such as statistical modeling, machine learning, and artificial intelligence to prognosticate. Today’s predictive analytics thrives on big data; the richness of structured and unstructured data sets allows for the creation of powerful predictive models.

Whether your company is just getting started with predictive analytics or you are launching the latest project within your predictive analytics initiative, there are a few things you should do before jumping in:

1. Needless to say, the first task is to clearly identify one or more business issues that you want to address. For example, increasing recurring revenue on a poorly performing established SaaS product.
2. Obviously, once a business issue has been identified as problematic or needing improvement, measurable success criteria need to be defined. For instance, reducing churn and increasing customer acquisition rates.

Planning for Project Success

Once the business needs have been identified and executive sponsorship established, project leaders should bring together a team of stakeholders across functional units/departments to plan out the project, including:

1. A decision-maker from the business or product line(s) who evaluates cost and business benefits to validate the business need for a project.
2. The head of the data team should help to quantify the effort and feasibility of project scalability as the project moves from sandbox to production.
3. The head of data science should be part of this to help identify data, as well as resource gaps.
4. IT is a key partner, and the head of IT and infrastructure needs to be involved to plan for production deployments.
5. The head of product management needs to be involved when the predictive models require integration into products (internal or external).
6. In light of ever-changing regulations, a security and governance expert should be part of the team.

Once established, the planning team becomes the steering committee and serves to move the project forward and eliminate roadblocks by determining project goals, setting expectations, identifying gaps, and opening communication channels.

Define Success

Clearly define the business issue, and identify and define success metrics. The steering committee should also be responsible for distinguishing secondary metrics and dependencies. For instance, you could establish a goal to improve recurring revenue by 5 percent, which may be impacted by external factors not influenced by predictive analytics. Break down the contribution of each action, such as customer acquisition rate and churn rate in relation to recurring revenue, to help determine whether progress aligns with your goals. While it is important to establish the business success criteria, it is also important to set infrastructure governance policies over the usage of resources over the lifecycle of each initiative.

Build Staffing plan

If this is a new initiative for the company, then it is necessary to have a short-term and long-term plan to build out the skill set required. Predictive analytics requires a combination of statistical modeling and computer programming skills along with sound business domain expertise. These are scarce skillsets and companies should have a plan in place to hire and retain talent.

Set the right expectations

Every predictive analytics model has a margin of error, and the steering committee will need to identify and quantify the risk involved with acting on an inaccurate prediction. Some business processes have a very low tolerance to risk, and the steering committee must decide how to drive business decisions with predictive analytics.

Identify data and technology gaps

Clearly identifying sources of data—internal and external—is imperative, as is quantifying the effort involved in pulling data from these sources. At the same time, the steering committee should also evaluate existing technology solutions, identify gaps, and make a recommendation to build or buy.

Establish channels of reporting and communication

Every new initiative needs continued executive sponsorship, and the steering committee must also clearly define a method and cadence to communicate progress to stakeholders and executives.

Once the business needs and goals have been defined, a typical best practice is to set up a nimble innovation team—consisting of data scientists, data engineers, product managers, application developers, and other domain experts—tasked with scoping out projects, prototyping, and testing solutions.

Bridging the Skills Gap

It is no secret that people with the skills required for predictive analytics such as data scientists, data engineers, and domain experts are in short supply. The success of an enterprise’s predictive analytics initiatives depends on acquiring the expertise to build, deploy, and maintain such projects. In the short term, hiring contractors and consultants to kick-start projects and get them off the ground is a viable option. However, for the long-term viability of predictive analytics projects, it is necessary to build effective hiring and training programs by:

1. Setting up an innovation center in a geographic region where data science talent exists to build out the initial prototypes and train the existing workforce.
2. Identifying high-potential employees, incentivizing and training them on the job.
3. Partnering with educational institutions to build a feeder program for entry-level candidates.

Predictive analytics should be part of a longer-term enterprise strategy. In the short term consultants and contractors can fill the short term talent gap, while internal skills are being built out.

Managing Change

The implementation of a predictive analytics project can result in significant changes to the way current employees work. Change management and training should be a part of any project not only for the success of the project but also to ensure that employees do not feel threatened and are capable of interpreting the results for various business scenarios.

To that extent, enterprises should ensure that employees are familiar with how data is collected, stored, and processed. Data users must understand—at a high level—how the model produces forward-looking metrics, the margin of error, and how to interpret results and incorporate them into business decisions. Enterprises should look to not only augment existing business processes with predictive analytics, but also identify opportunities to completely revamp business processes with predictive analytics, and retrain employees appropriately.

Investing for the Long Term

Predictive analytics have powerful applications in today’s business environment. However, enterprises that seek to move beyond experimentation and become truly data-driven will need to train their employees and provide controlled self-serve access to relevant data. This self-service strategy is only effective when employees know how to test out hypotheses and propose solutions that will improve business outcomes. When in production, governance policies need to be in place to ensure the continued accuracy of models. As production deployments scale, costs should be closely monitored and optimized on an ongoing basis.