Conquering the Top 5 Misconceptions About Predictive Analytics

Predictive analytics is often cited as a key business driver for Big Data. It is easy to see how predictive analytics, when done right, can represent a strong competitive advantage for companies. Such as:

• Enabling companies to be proactive which translates to reduced cost, improved service and/or being the first to market
• Making customers feel you’ve read their mind which improves customer loyalty and retention

So let’s discuss the common misconceptions about predictive analytics and how to overcome them:

1. We need to start small and iterate, so let’s start with a limited feature set.

A key roadblock for implementing analytics projects is the availability of quality data.  In the initial phases of analytics projects, obtaining a wide feature set (a.k.a. attributes) on a reasonable number of data points is more important than having a limited feature set on a large number of data points.

Why? Because starting off with a sizable feature set helps us triangulate the data better, identify the key features that most reliably drive the predictions and cull out the others. So what does it mean to start small in analytics? It means:

• Focusing on a limited set of the most pressing business problems and associated use cases
• Using a representative sample of data points as input into the analytics (as opposed to trying to acquire all the data points)
• Applying the business actions to a limited set of instances to minimize business risk and build a feedback loop for iterative model refinement

For example, when modeling customer behavior and personalizing customer experience, it is better to consider all attributes of a representative set of customers, rather than a limited number of attributes of the whole customer base. Similarly, you may want to personalize experience for a subset of customers and iteratively refine the personalization toward the larger customer base.

2. We are predicting outcomes reliably, so we are done.

Generating meaningful insights, as difficult as it is, is easier than acting on them. After all, insights have no business value unless acted upon. Why is it more difficult to act on the insights? Because the opportunity cost of not acting, the cost of acting on incorrect insights and/or the unforeseen side effects of action have to be well understood.

So how do we make the process of acting on the insights less risky? By starting small and iterating quickly. Recall how Starbucks introduced oatmeal breakfasts in just a few carefully picked stores before expanding to a large market?  This is a technique also used routinely in the retail industry.

3. We are successfully taking actions on our predictions, so we are done.

Predictive models require constant care. All analytic models start off with environmental assumptions. These assumptions need to be constantly re-evaluated and the models need to be updated as the environmental factors change.

Lacking this, the models degrade over time and lose their predictive power. While lifecycle management of enterprise software protects existing enterprise assets from degrading in business value, lifecycle management of predictive models also preserves customer loyalty, competitive edge and revenue. Budgeting for the lifecycle management of predictive models is key when embarking on Big Data projects.

4. Let’s minimize our assumptions and let powerful analytics algorithms do the heavy lifting.

All analytic models start off with environmental assumptions that need to be adjusted as part of model life cycle management. Some analytics practitioners argue that by minimizing the explicit environmental assumptions and providing the model with all possible features, the model can self learn using powerful algorithms, thereby minimizing or even eliminating lifecycle management costs.

The practical reality is that explicit assumptions bring domain specific knowledge to bear, consequently reducing the search space and increasing the efficiency and accuracy of the model.

For example, the simple assumption that mobile device usage tends to taper off for business users after business hours (unless they are BYOD users) can help reduce search space in customer segmentation. There is a cost/performance balance between how many assumptions you want to maintain for a model vs. how much self learning you want the model to do. It is important to understand this balance to receive a good return on investment on predictive analytics.

5. Data does not lie so we need to act on what the data tells us.

Business decisions have always been made using business intuition. With predictive analytics generating new data-driven recommendations, the question becomes “how do you reconcile the two?”

While data does not lie, there may be multiple factors that affect the quality of data-driven insights: poor quality of source data, environmental assumptions that are off target, errors in the approach of building predictive models, etc.

On the other hand, business intuition is more mature because it has been fine tuned over time based on the feedback loop of business performance. So the real contribution of predictive analytics is not to substitute business decision making, but to help evolve from “gut feel-based” decision making to “evidence-based” decision making.

Case in point, we all know that loyal customers buy on a regular basis, but when do we need to reach out to customers who may be at the verge of buying? Evidence-based quantitative backing could be invaluable when deciding how to manage these customers.

To round things out for this post, predictive analytics, which was once the prerogative of data scientists is going mainstream thanks to increased awareness and tools that abstract out the task of complex analytics design. So it is important for the new analytics practitioners to be aware of the misconceptions around predictive analytics and work around them to help make meaningful business impact.