This article on predictive analytics is the fourth in a series of guest posts written by Dan Vesset, Group Vice President of the Analytics and Information Management market research and advisory practice at IDC.
Analytics solutions ultimately aim to provide better decision support — so that humans can make better decisions augmented by relevant information. Decision support capabilities can be segmented into five related categories, each of which is deployed to answer different types of questions:
- Planning analytics: What is our plan?
- Descriptive analytics: What happened?
- Diagnostic analytics: Why did it happen?
- Predictive analytics: What will happen next?
- Prescriptive analytics: What should be done about it?
In this series of blog posts, we’ll address each of these analytics capabilities. For a fuller introduction to the topic as a whole, see the first post in the series. This fourth post focuses on predictive analytics.
Predictive analytics: What will happen?
As your enterprise is managed to a set plan, and descriptive and diagnostic analytics are used to measure execution against the plan and understand reasons for deviations from it, there is an opportunity to start predicting what is likely to happen next. Being able to effectively apply predictive analytics creates an opportunity for targeted resource allocation that is not driven simply by experience or a “gut feeling.”
A prediction will always remain a prediction — with its inherent shortcomings in knowing the future. However, as the British philosopher Carveth Read said, “It is better to be vaguely right than exactly wrong.” In today’s world of big data, a growing number of enterprises are deploying powerful predictive analytics to increase their chances of being right (and often much more than vaguely so). Enterprises are using predictive analytics to determine the next best actions to take by sales or customer service reps, which physical assets are likeliest to malfunction, or which candidates to hire.
Predictive analytics are the domain of data scientists, who are tasked with steps in the analytic workflow represented by the following five categories:
- Identify business outcomes: Determine what questions need to be answered using predictive analytics. If correct outcomes are not identified, running predictive analytics is like throwing darts in the dark. Also identify drivers (independent variables) that will most likely impact the outcome (dependent variable).
- Determine data required to train: Predictive analytics require data from multiple sources, so analysts must identify current data sources. If existing sources are insufficient, they must acquire data from other sources to ensure that models can be accurately trained.
- Determine types of analysis: Different techniques are suited to answer different questions depending on the amount and type of data available. Statistical analysis, neural networks, machine learning, and data mining are all examples of sophisticated techniques that can predict outcomes.
- Validate results: Advanced analytics cannot be used like a black box; bad training data, incorrect algorithms, and poor assumptions are just some of the pitfalls that can result in false predictions. Data scientists need to work closely with line-of-business analysts and leaders to ensure that the predictive models make business sense.
- Test predictions on performance: Predictive models need to be continuously tuned to improve accuracy. If a model fails, analysts must identify the root cause and retrain and test to improve the models.
Modern predictive analytics solutions must provide data scientists with a productivity workbench that supports all of these functions. But predictive analytics isn’t only about data scientists’ productivity. The value of predictive analytics comes from their deployment within other applications in the flow of business. That means a predictive analytics solution must support execution of predictive scores at scale.
It must also support model management. Enterprises should treat their predictive models like any other asset or intellectual property. Any organization that uses predictive analytics will have a portfolio of models that needs to be governed over time. As data and business contexts change, some models will deteriorate and need to be re-calibrated.
Finally, predictive analytics are a necessary step toward prescriptive analytics.
For IBM’s view on the Analytics Cycle, check out our smartpaper, “How Can You Trust Your Data Without the Big Picture?”