Fantasy Football trades: How IBM Granite foundation models drive personalized explainability for millions

With almost 1,700 players in 272 games, the amount of data generated during the NFL football season is enormous. Fantasy football team owners are faced with complex decisions and an ocean of information. Deciding who to start, who to bench and who to trade each week can be a daunting task. It can also be a lot of fun—and that’s why the ESPN Fantasy app engages 12 million fantasy football users each year. 

For the last 8 years, IBM has worked closely with ESPN to infuse its fantasy football experience with insights that help fantasy owners of all skill levels make more informed decisions. These insights take the form of player grades that help end users find the best players to trade or pick up from the waiver wire. Andthis year, the team is going even deeper, adding a new feature that unpacks the reasoning behind the AI-generated grades. When a user taps on a player to acquire or trade, a list of “Top Contributing Factors” now appears alongside the numerical grade, providing team managers with personalized explainability in natural language generated by the IBM® Granite™ large language model (LLM).  

As in real-world organizations, managers of fantasy football teams need clarity about the “why” behind AI-generated output. “Explainability—the reasoning behind the output—is becoming almost as important as the output itself,” says Aaron Baughman, IBM Fellow, Master Inventor and IBM Quantum™ Ambassador.  

The Top Contributing Factors provide explanations based not just on a player’s raw performance, but also the specific ways in which they will complement your fantasy roster. Here’s how it works. 

Making the grades 

As players become available in your fantasy league’s weekly waiver wire, they’re given a personalized Waiver Grade that takes into account the strengths and weaknesses of your roster, as well as the rosters of other teams in your league. The grades even incorporate the custom settings of each league. The score is based on the value they would add to your team, compared to the average grade of players in the same position in your current roster. Trade Grades work similarly, based on the relative benefit a player from an opponent’s roster would add to your team. 

This process begins by calculating raw grades using a rules-based system in combination with a plurality of machine learning models. “Working with our developers and football experts from ESPN, we determine grades based on a variety of factors,” says Aaron Baughman. “How many leagues own this player, what percentage of leagues start the player, what are their projected seasonal stats, who have they played, who will they play—these types of predictors are rolled together into a single score.”  

The grading system is written in Node.js and Python, supported by a scaled-out workflow that analyzes the billions of data points generated over the NFL season. The results are then saved to the cloud. 

Personalized analysis at scale 

Closer to the consumer, a Node.js “team needs” application personalizes these grades specific to a user’s team every 10 minutes. “The team needs application considers your roster, your league and its specific settings based on an algorithm we designed that was over a year in development,” says Baughman.  

Why did it take so long? In a word: scale. There are 12 million or more fantasy users every week, and some  days—generally Tuesdays and Thursdays—can have significant usage. This can result in the app receiving thousands of hits per second, which are scaled out across pods on a Red Hat® OpenShift® cluster. 

When a team owner taps on a waiver wire player in the application, the program runs and provides a grade customized to their team. “Although points per reception (PPR) is the most common scoring system, leagues can have an infinite number of custom settings. As a result, we couldn’t pre-compute any grades, they had to be personalized,” says Baughman.  

Beyond fantasy football, one can imagine many business use cases that could benefit from this combination of high-scale data analysis and personalization. 

Personalized explainability at scale 

As in any management situation, it’s important to summarize analytical insights concisely for busy decision-makers. That’s why the system winnows down these insights to an executive summary of three Top Contributing Factors. 

“Working at this scale was a challenge, so we invented a new algorithm,” says Baughman. “In watsonx™, a Granite generative AI model outputs fill-in-the-blank sentences. At the edge, these incomplete sentences are then personalized based on value percentiles.” The potential adjectives and phrases change across numerical values, for example from “no help at all” to “greatly improves” (your position).  

It’s a multidimensional model with 12 types of contributing factors, and hundreds of permutations. A few of the most prevalent factors are percentage of teams owning a player, percentage of teams starting a player, and score projection. Inferencing is run approximately every two hours and joined with raw grade scores to fill in the blank. 

“For Trade Grades, when you click on an opponent’s team we surface three positions your team needs to address, such as tight end, wide receiver or  defense—the circle turns green if it’s one of your needs,” says Baughman. “I can also see my opponent’s needs, which encourages fair trades.” 

Subsequently tapping a player card reveals the overall trade grade, derived from the combination of raw analysis and team needs, with the three top contributing factors. 

Driving business insights and explainability 

Whether you manage a fantasy football team, a company or a business function, the IBM watsonx™ AI platform can help you make better decisions. It helps you collect, store and analyze enterprise data relevant to your use case, then use a variety of machine learning and traditional AI models to evaluate strengths, weaknesses and  opportunities—and deliver timely, up-to-date insights with fine-grained contextual detail. Enterprise-focused Granite LLMs help provide natural language explainability, along with a host of customer service functions. Add the scalability of Red Hat® OpenShift® on hybrid cloud and look forward to your organization having its strongest season yet. 

Explore the watsonx AI platform

Redefine how you work with IBM AI consulting services