Wouldn't it be great if you could see the future? People with decision-making responsibilities can get stressed out trying to predict the downstream results of their choices. That's the thinking behind predictive analysis, a method for using historical data to predict the outcomes of present decisions and project them into the future.
While it's not totally foolproof, predictive analysis can give you quite an edge when you're working with data from historical inputs to spot trends and choose the right path. You can use this toolkit in all kinds of areas, from marketing to hiring to brand performance over time.
This article explains what predictive analysis is, how it works, and where you can use it to fix thorny management problems. By the end, you should be comfortable with the concept and ready to start working predictive models into your normal decision trees. You should also finish this piece with some great ideas about how you can use the predictive analysis toolkit to make your work much more productive and less stressful.
What Is Predictive Analytics?
Predictive analytics is a set of tools for modeling future performance based on past performance. It's a set of mathematical processes for projecting trends outward from the data you have to the outcomes you're trying to predict. By picking the relevant historical data and applying an appropriate set of tools, you can reliably predict the outcomes of policies and initiatives you're contemplating now.
To put that in real terms, imagine that you're an HR manager charged with hiring people to write code for your company's IT department. In your experience, certain colleges produce graduates who do better with this kind of work. So you screen your applicants for keywords like "CalTech," "MIT," or "University of California Berkeley." In a very simple way, you've just applied your own predictive analysis to make a meaningful connection between an applicant's school and how well they might perform at your company.
You'll need some hefty computing power and good machine-learning algorithms to handle more complex associations. Say your past sales data says your company sells more fireworks in the weeks leading up to New Year's and the Fourth of July, but your subsidiary sells lots of turkey in November, right before Thanksgiving. In another fairly obvious association, you can predict in advance how your fireworks and turkey sales will look just from past performance and the calendar date.
Taking It Up a Notch
Going further, you may have experience launching brands on social media, but your past performance has been sketchy and disappointing. Using marketing tools with advanced predictive analytics models, your data analysis shows that brand messages travel well from Twitter to Instagram, but they hardly ever go in the opposite direction. In this case, you may develop actionable insights like always starting a campaign on Twitter or building out your Instagram following to achieve better performance.
So far we've seen a few very simple ways to predict future events using some exceedingly simple predictive analysis methods, but it can go way up from here. Your FICO credit score is a kind of predictive analysis. By looking over your history of on-time payments, available credit, current credit utilization, and several other factors, the FICO score predicts with a high degree of accuracy how likely you are to default on your debts. It can even factor in the number of credit checks you've had in the last 2 years to determine whether you're desperately looking for credit extensions, a sign of trouble ahead.
Back in the workplace, you might be a QA officer whose job it is to keep the machines running. Any downtime could cost your company money, so you have to predict when and where the breakdowns will likely happen. Your machine learning reviews years of past data and determines that breakdowns occur when a critical part wears out, which happens faster when the operating temperature is high. The actionable insight here is to run the machines at lower intensity or to invest in a bigger cooling system to keep temperatures down, reduce wear on critical parts, and extend the life of the machines, potentially saving millions of dollars.
Predictive Analytics Vs. Data Analytics
Now that you have a broad-strokes picture of what predictive analytics is, here's what it isn't. Or rather, here is the dividing line between predictive analytics and other, related concepts that it can easily be mixed up with.
First, predictive analytics isn't the same as the more broad data analytics. It would be accurate to say that it's a subset of data analytics, one that focuses on a limited set of tools to build out future-oriented models. The other types of data analytics are:
- Descriptive analytics: An accurate and rationally organized accounting of what has happened in the past, presented in an intelligible way
- Diagnostic analytics: Figuring out why things happened the way they did in the past, ideally to gain insight for later predictive models
- Prescriptive analytics: Actionable insights into where you should go from here, and a set of instructions for what (and what not) to do while you're working toward a goal
There really aren't bright, clear lines separating these types of data analytics, and every one of them can inform the others. A really powerful approach to analytics uses all of them in their proper place to build robust models of understanding for complex affairs.
For instance, you might be an engineer trying to design a safer car. You'll need tons and tons of descriptive analytical data from past crashes, including injury reports, to build out a model for how cars break. The picture you get from this might reveal that crash victims frequently get neck injuries. This is a description of raw data, presented in a way you can read, but you need to go further to make sense of it.
So you know there are a lot of neck injuries in car crashes. Your next step, diagnostic analysis, tells you why. Apparently a lot of people bang their heads on windshields or whiplash back and forth. Sometimes, a heavy object hits them from behind during a crash, which causes the injuries you're seeing.
In the third step, prescriptive analytics, you develop a number of suggestions for how to prevent neck injuries in car crashes.
You might consider designing the car so it won't move until the driver is wearing a neck brace. You might build a safety cage around the driver's head made out of heavy lead shielding. You could install an automatic ejector seat that senses problems and vaults the driver out at the first sign of trouble. And so on. No bad ideas, just ideas at this stage.
Your job really comes together in the predictive analysis stage here. Running past performance models, you learn that car owners get annoyed at intrusive safety devices and will likely disable the neck braces, which makes them a nonstarter. Adding weight in shielding and reducing drivers' fields of vision are counterproductive. Past data from the aerospace industry tells you that automatic ejector seats are generally a bad idea.
But there are a few promising ideas in the pile, which your predictive analysis reveals:
- Seatbelts are a cheap way of keeping people steady in a crash.
- Airbags cushion moving bodies.
- Safety glass breaks easily, reducing head and neck injuries.
- Collision avoidance reduces the risk of a crash in the first place.
So you build these reasonable ideas into your upgraded design, which your analysis predicts will get the best results for the money spent. The cycle then repeats, as a new generation of safer cars go forth, get in crashes, and generate more data to work with, making cars ever safer in the future.
Common Types of Predictive Models
All of this is great so far, but what actually are the tools of predictive analytics? How does it work? What do the artificial intelligence, machine learning, and assorted whatnots actually do?
Well, there are four basic models for this kind of analytics. Different platforms use all or some of them, and their use is really context-dependent. Here's a brief rundown on the different types.
Classification could just as easily be called segmentation. This approach breaks large data sets into chunks with meaningful relationships. If you've ever gone near a market research project, you're familiar with demographic segmentation. If you've tried marketing skateboards to seniors, you've already learned how important some classifications can be.
Specific models for accurate classification include logistic regression analysis, decision trees, random forest, neural networks, and Naïve Bayes analysis.
Forecast is one of the most commonly used approaches, party because it's so versatile and partly for its ability to put real-time number values to complex, numberless historical phenomena.
For example, you can use the history of product launches under your brand to predict web traffic on the day you launch a campaign. Using this number, you alert your web host and ask for extra bandwidth. As a result, you get tons of traffic, your site doesn't crash, and sales are strong.
Clustering models are really powerful, and the rise of deep-learning algorithms has made them scary effective. These work by spotting correlations between things in the real world, comparing them with other things, and then making predictions based on associations.
For example, your clustering model might notice that people who buy diapers, baby formula, nursery furniture, and high chairs are really likely to also buy healthcare insurance and school supplies in the near future. Storing this information away, it can become the basis of a successful sales pitch targeted at these individuals as much as 18 years in the future, when a client is marketing student loans.
As you might have guessed, a lot of these models exist. Common clustering algorithms include k-means clustering, mean-shift clustering, density-based spatial clustering of applications with noise (DBSCAN), expectation-maximization (EM) clustering using Gaussian Mixture Models (GMM), and hierarchical clustering.
Time Series Models
Arguably the first time series analysts were the hunter-gatherers who noticed animals migrated with the seasons. This approach follows cycles to predict future events.
The example from above about turkey sales going up in November is an example of time series analysis. Specific model types you could use here include autoregressive (AR), moving average (MA), ARMA, and ARIMA.
3 Benefits of Predictive Modeling
There's a lot to be said for these predictive models and how they help with your decision making. For a clearer picture, here are three places you can apply the models to get great results.
Security and Risk Reduction
Security is mostly the business of predicting people's behaviors and accounting for them in advance. Locks on your doors predict that people will try to turn the doorknobs at some point. Forensic audit trails anticipate someone will try to embezzle money.
In marketing, your historical data might suggest that information leaks are most common with third-party vendors. So you decide to hold information back until product launch dates or even to segment information to keep it secure.
Predictive tools can vastly improve your operating efficiency by helping you forecast demand. If you have a call center that's really busy in the early morning and late afternoon, but it's a ghost town at all other times, you have data that helps you schedule shifts.
Bringing your hourly staff in during peak hours, then going slack when nobody is calling, puts your resources right where they're needed at the right times, rather than being shorthanded during rush hours and overstaffed when it's quiet.
Improved Decision Making
This is the big one. You're turning to predictive analytics in effect to get a cheat code for future performance. Think FICO scores again. Should you extend credit to a bad credit risk with a low score? Should you refuse to send out a new credit card to someone with a score of 790? The analytic model helps you predict future outcomes based on past performance.
5 Ways to Use Predictive Analytics to Your Advantage
Customer and Audience Segmentation
You already know your customers are different from each other, but they tend to cluster in groups that big data sets reveal. Data mining has a bad name, but it's the first step to building out sensible clusters you can generalize about and target more efficiently.
New Customer Acquisition
Winning new customers is great, and the ability to do it reliably with your marketing campaigns gives your brand a competitive advantage. Techniques such as linear regression models help you hone your workflows and get the most out of new customer acquisition campaigns.
Sorting the sheep from the goats is a big part of business intelligence, especially with lead scoring and developing a new following. Propensity modeling shows where the public is likely to go with you and which prospects are the hottest.
Content and Ad Recommendations
Ads can be annoying if they're unwanted or if they advertise products the viewer doesn't need. Collaborative filtering screens out the annoyance and makes your ad messages far more likely to land.
Personalizing Customer Experiences
Whether you're a retailer, in the financial services field, or doing weather forecasting for shipping companies, automated segmentation helps you sift through all the outputs you could be generating and deliver a customized experience for everyone you interact with. This is optimization on a personal level, and it has the potential to make the strongest impact in your marketing since it connects on a personal level with everyone your brand interacts with.
Moving Into The Future
That was a lot to go over, but hopefully you're now much better informed about predictive analytics tools and the awesome ways you can use them. There's also much more information available if you're looking for it. Subscribe to The CMO newsletter to stay on top of the latest tips, trends, software recommendations and more in marketing.