The Analytics Mindset - How to solve problems through modeling and data

When starting on a data project, there is a lot of technical work to manage the data and build a model. However, there is also a general scientific mindset that analytics professionals need to have in order to deliver maximum benefit to their stakeholders. There are many people in the world who have learned this approach, but generally through osmosis or a natural disposition. My goal in this article is to explicitly teach the "Analytics Mindset" in hopes that more people can use it to achieve the things they want with the toolset of analytics.

The first step of the Analytics Mindset is simple to say, simple to share, and sometimes hard to keep in mind in the heat of a project. That first step is to:

• Focus on the goal.

In the domain of Operations Research (started during World War 2) we call this the “Objective.” The first thing you write down for a problem is what function are you maximizing or minimizing. Do you want to maximize expected profit? Minimize customer churn? Minimize the average wait time for service? Maximize customer engagement? All of these goals can be articulated by organization leaders who are in a position to guess how improvement on a certain priority will have the biggest impact on the business. Once you have a goal in mind, you can formulate a way to quantify that goal and measure it.

Once you have identified the goal, the next step in the Analytics Mindset is to understand what else your solution will need to do. This can be simply described as:

• Understand the constraints.

Unlike the first step, where you can reasonably expect leadership to give you a single goal and stick with it: navigating “everything else that this solution needs to understand” is not a straightforward task. Machine learning and optimization algorithms are excellent at finding the perverse solutions that meet your quantified objective, but entirely miss the point of the exercise. An algorithm that is asked to maximize profit might get very excited to charge an infinite number of customers infinite dollars each for a product. We often rely on people to make these kinds of judgment calls, and people are usually very understanding audiences (the link is to the exact instruction challenge for making a PB&J sandwich which you may have seen as an introduction to the difficulty of programming). But in the case where we are asking an algorithm to solve a problem, we must identify all the properties that would make a solution useful or not. Oftentimes, this is an iterative process of showing stakeholders the solution the algorithm found, and asking them what we are missing in the constraints so far.

After you have identified both the goal and the key constraints, you have the basic framework for solving the problem. However, as analytics professionals, the typical next step is to bring in the actual data. This is made up of two key steps:

• Work through how uncertainties throughout your models and data might impact how well you achieve your goal.
• Work through how uncertainties throughout your models and data might impact how well you satisfy your constraints.

As you dig into the data, there is a true “rubber meets the road” kind of moment as you see which information you have at all, and which information has important limitations. This is also when you need to dig into which models can help solve the problem at hand. Some of your models may be focused on addressing the limitations of the data. Other models will be about representing how the input decisions and data will relate to the outcome variables. Because the world is complex, someone with this Analytics Mindset will be trying to figure out how to mitigate the approximations and limitations in the data and models as they relate to the goal and constraints. For example, a company that knows their market size but not their market share might realize that investing in third-party-data will significantly improve their confidence. From the reverse side, the Analytics Mindset can also help identify previously unknown limitations of the data and models. 

Often in parallel with the task of modeling through available data, analytics professionals often will identify issues that are both key and highly uncertain. As these concerns rise to the surface, the final skill of this mindset is on quantifying tradeoffs.

• Because your model and data will never be perfect, consider how to represent tradeoffs in your solution to be able to have decisions that make sense even when some of the uncertainties are realized.

This element of the Analytics Mindset is the hardest to quantify, but is also perhaps the single biggest driver of adoption for analytics solutions. Users will often hold back from following the analytics because of doubts in their mind that it covers enough of reality to be trusted. Those concerns are often valid, but those without this framing do not have the words to articulate why they are not following the analytics. A project that is being led by someone with this mindset will be able to work with stakeholders to understand and incorporate their concerns. Allowing the solution to incorporate those major drivers of uncertainty and still provide a “best” recommendation allows the end users to quantify the costs of uncertainty and the benefits of a particular decision.

I hope this article has helped introduce you to a new way to think about solving problems with analytics. For more information on this approach, I highly recommend joining INFORMS as well as trying to think about the meta-problem as you build any analytics solution you hope people will actually use.