Walter Sun - PhD, Vice President of Applied AI, Microsoft’s Business Applications & Platform Group
Every year, millions of basketball fans fill out brackets for the Division I Men’s College Basketball Tournament hoping to predict the team that will win it all. The Microsoft Business Applications Applied AI group has joined in on the fun, leveraging AI to help guide our bracket picks.
So far, we’ve had a pretty good run. Out of the 17 million brackets submitted for the 2022 ESPN Men’s Tournament Challenge, the Microsoft Business Applications Applied AI group’s bracket is ranked 21st on the leaderboard, which after four rounds (semifinal stage) is more accurate than 99.9997 percent of all brackets submitted at the beginning of the competition.
In this post, I’ll discuss how we built this bracket and explain how we use similar techniques to power AI capabilities in Microsoft Dynamics 365 applications. Why does this matter? Because businesses want to make better decisions and achieve better outcomes, and—as the accuracy of our tournament predictions help show—turning to AI to provide those insights can deliver impressive outcomes.
How we build our tournament AI model
Since readers are probably more interested in the construction of our bracket, let’s start with that (for business-minded readers, you can skip to the next section). We’ve previously explained how to build better brackets, and we had deeper dives providing details of our models. To summarize, we take over 10 years of historical data for teams and analyze many factors, including:
- Regular season and tournament outcomes.
- Current player composition, excluding injured team members unavailable for the tournament.
- Historical competitive success of the head coaches coaching this year’s teams.
- Success away from home and in neutral sites, which is where all tournament games happen for the men.
- Historical analyses of seeds and conference successes.
With these factors, we created a model which determined that 5th seeded Houston was dramatically underrated at that position and we were thus able to correctly predict their upset wins over 4-seed Illinois and 1-seed Arizona and that 8th seeded North Carolina would upset 1-seed Baylor and 4-seed UCLA. The model understood team strengths well enough that by the quarterfinals, each of the four teams we predicted to make the semifinals was the stronger remaining seeded team in the four contests. And fortunately for us, all four teams delivered to ensure we’d have a perfect semifinal bracket going into this weekend.
Solving business problems with AI
Businesses have many questions they need answers to today. For example, what’s the likelihood that a customer will churn? What’s the predicted lifetime value of a customer? What products should be recommended to a customer? Will I have enough supply of a given product on the shelf? Accurately answering these questions leads to better outcomes, as customers are retained, loyalty programs are designed to reward your best customers, and customers see the products that are most relevant to them available on the shelf. We have built easy-to-use, out-of-box AI models that are trained on your data so you can get the predictions optimized for your business to answer these questions.
These models use the same concepts that we used to create our bracket, as the propensity of a business outcome can be modeled similarly to that of a sports outcome. Take, for instance, the classical Recency, Frequency, and Monetary (RFM) model for predicting churn.
- For recency, in the basketball model we look at streaks and “last 10” performance, while in business churn, we look at the recent purchase history of a customer.
- For frequency, we might analyze the number of wins a team has for basketball, while analogously we look at the quantity of purchases from a customer for our business applications models.
- Finally, for monetary, this would be the point differential in basketball and quantity of spend in our churn model.
This is only a subset of all features we analyzed to give you an idea of our approach.
Beside the questions above, there are a lot of other questions that businesses are asking in every domain. For example, what are customers saying about our product? How do we identify business entities in text? How do we find the most relevant news articles about suppliers? How do we analyze your processes and automatically label the activities that are being done to identify bottlenecks and suggest improvements? We have AI capabilities to help answer these questions available in our Dynamics 365 product family today.
Of course, AI models need not be fully self-service as human intuition and guidance can be included when desired. This is why we have spent time discussing the importance of explain-ability of models in a recent post. People can take AI and use it as a guide to improve decision-making both in creating a bracket as well as solving your business problems. Namely, our AI models explain the insights so that you can add in your own intuition when planning your next marketing campaign or building your next material requirements plan.
Keeping pace on and off the court
The dynamic nature of our models can be seen with the tournament, too, where we incorporated the early results from additional post-season college tournaments to adjust the modeled strengths of conferences before main bracket play began—in 2022, this meant observing a stronger Atlantic Coast Conference based on Virginia upsetting Mississippi State and Wake Forest winning in outcomes from another tournament’s first round. Such a signal increased the model’s confidence in this conference, leading it to correctly pick a Duke versus North Carolina semifinal and a quarterfinal run for Miami.
In business, the only constant is change. How do you pivot and adjust to the latest disruptions in your supply chain? How would you route customer support calls to the best agents based on real-time staffing changes and that day’s routing behavior? We’re running a full court press in solving these types of problems, so if you want to learn more about how AI can help your business, try our line of business applications, visit our Insights blog, and come apply to the many AI-powered jobs we have for Dynamics 365.