This month, our team read the book Predictive Analytics by Eric Siegel, and got together to discuss some of its key concepts. We split into small groups and completed an activity in which we walked through potential analytics projects for companies and organizations in a variety of different industries—brainstorming predictive analytics (PA) use cases, which data sources each organization could utilize to create models, who at the organization would be involved, and what specific actions could be taken for each use case. Our VP of Predictive Analytics, David Cieslak, also gave an interesting talk about machine learning, predictive analytics, and the strategies he has developed while working with our clients.
Since this book got us thinking and talking, we thought we should share some of the major takeaways with our blog readers:
With Prediction, a Little Goes a Long Way
Traditionally, businesses have used forecasting to predict overall trends on a macro level, but this can only go so far. With the large amount of data being collected and available for analysis on ever-more-powerful computer systems, we can make predictions on an individual level—who is most likely to “consume, think, work, quit, vote, love, procreate, divorce,” and more. This allows a company to target specific individuals within a population, instead of trying to reach everyone and hope for the best. Although no prediction is 100% accurate, a model based on data improves upon a random (or even educated) guess. Businesses realize that even a small improvement can make a difference, and they are taking action; PA can help inform and improve those actions.
Action Is a Key Component of Predictive Analytics
Implementing a PA project isn’t as easy as flipping a switch. There are a number of considerations that must be addressed. First and foremost, being able to predict something doesn’t do you any good unless you are able to act upon those predictions. Any PA project must seek to answer an “actionable” question. In other words, the answer must lead to an action.
Good Models Take Time
Another consideration is that analytics can’t be completed instantaneously. It takes time to properly organize the data, develop a winning model, and test to ensure the model accurately predicts what it seeks to predict. To do this, data scientists employ a technique called “backtesting,” in which they run the model to predict past events before using the model to predict the future. This confirms the model is giving the answers we think it should; it’s a mathematical “logic check.”
Privacy Concerns Must Be Properly Addressed
Data is a valuable resource, with the potential to make a big impact in many realms. But this also endows purveyors of data with a number of responsibilities. The most pressing concern is privacy—considerations must be taken to be sure that data is used responsibly, and that the actions taken on predictions respect the individual’s right to privacy.
One may mistakenly assume that access to all this data allows data scientists to see too much about an individual, but it is quite the opposite. Data scientists aggregate data from many individuals in order to discover subtle patterns that only exist across large quantities of data. When applied to any given individual, all predictions must be taken with a grain of salt. Each individual is different, and has their own free will. Care must be taken to remember that just because an individual is predicted to do one thing, does not mean they will necessarily do what was predicted.
Organizations should establish and discuss their data management strategy, deciding: what is stored; how long to retain; who has access; where to share; what to merge; when to react; and finally, to what end.
Simply Exploring Data Can Lead to Insights
The instrumentation of everything—from cars, to phones, to watches, and much more—is causing data to be created in unprecedented volumes. Not only is the total amount of data increasing, but now the data collected also reflects behaviors rather than mere demographics. When data is analyzed at large volumes, unseen and surprising insights emerge. Since everything is connected to everything else, the past data tells a story—and data is always predictive. This is referred to as the Data Effect. However, it must be noted that patterns and correlations do not reflect a cause-and-effect relationship—and they don’t have to! The knowledge of a relationship is enough to make a prediction. It is also worth noting that insights can even emerge through the data-preparation phase. Simply exploring your data can lead to interesting insights and make an impact on your business.
Decision Trees Are Powerful and Effective
Using data to inform business decisions is not a new phenomenon—in particular, banks and insurance agencies have been using data to manage micro-risks such as loan or insurance policy approval for years. However, with the advent of big data analytics, every industry can benefit from the ability to use data to generate predictions. But how does a model work?
The decision tree is a popular method for modeling because it is both effective and relatively simple to implement and understand. It’s also a familiar concept to most people; it works much like a flow chart. A decision tree works by separating a population by a certain measure, such as high-risk vs. low-risk. It uses multiple variables to further segment each group; and at the end, calculates a predictive score. As a decision tree “grows” downward it defines and refines smaller and smaller segments, essentially squeezing knowledge from the data like a fruit juicer. Data scientists must be careful not to grow the tree too large, or they risk mistaking “noise” for useful information; this is called “overfitting.”
In order to prevent this fatal design mistake, predictive analytics must occur at the intersection of art and science. By using both “artistic design” (to avoid bias) and scientific measures (to evaluate effectiveness), a data scientist can create an accurate, useful predictive model.
Ensemble Models Can Improve Upon a Single Model
There is not one, singular way to create a great predictive model. This is why creative solutions are imperative to data science. When Netflix, a movie rental and streaming company, challenged the tech community to create a better model for predicting movie recommendations, a myriad of solutions emerged. However, no single model was able to reach the goal of a 10% improvement upon the existing Netflix algorithm…until participants thought creatively and began to merge their models. It was risky for the teams to merge their models because each would lose their individual competitive edge. But there is wisdom in crowds and in multiple models. These meta-models employ what is called an Ensemble approach. An ensemble model increases diversity and overcomes any single model’s shortcomings, without adding complexity (overfitting), and can boost performance anywhere between 5-30% over a single model. Ensemble modeling is taking the PA industry by storm.
Uplift Modeling Optimizes Marketing Efforts
Predictive modeling does a great job of predicting who will “click, buy, vote, etc.” allowing businesses to take action on the resulting target segment. However, there are two pieces of information traditional predictions cannot tell you: 1) Who is likely to take the desired action despite NOT having received any outside contact (such as the person who is going to buy your product regardless of the advertising you sent them); and 2) Who would be dissuaded from taking the desired action if contacted? An example of this is the subscriber who is reminded by your mailed advertisement that their contract is about to expire, and uses that as a trigger to switch to a competing company. In this case, it is actually detrimental to contact these individuals.
The influence of an action on a person cannot be observed directly; you can’t simultaneously contact and not contact the same person. However, this influence can be inferred through predictive analytics. This process is called “uplift modeling.” In order to do this, data scientists compare the results of two separate data sets; one that was contacted, and one that was not. This is similar to any scientific study where one group is the “control.” In this way, it can be inferred who will be affected by a marketing attempt; and then a company can focus on converting or retaining those individuals. Uplift modeling offers the benefit of contacting fewer people while improving overall results. Imagine only contacting 60% of your targets and improving results by 36%. This is the “utterly fundamental targeting of marketing” and the promise of predictive analytics.
As you can see, this book contains many good nuggets of information. If you are interested in learning more about predictive analytics, its benefits and challenges, and examples of PA in action, this book is a great read that is suitable for a wide audience.