ESPN's subsequent fantasy performance predictions might then subtract a point or two from that average if that quarterback's next game was against a tough defense — as also measured by structured data. However, the network wasn't doing much to incorporate unstructured data. What were the beat writers who covered the quarterback's team on a regular basis seeing? What were the hardcore football analysts, who can better account for things like injury and off-the-field distractions, saying about the quarterback's matchup with a particular set of defensive backs?
Training Watson to make useful sense of that information was, as Baughman says, "extremely complex." First, IBM had to teach Watson to understand football by having the system "read" over six million articles about the sport. Next came teaching Watson how to understand fantasy football, which Baughman says required several IBMers annotating thousands of sentences and phrases so Watson could learn the game's semantics.
Baughman then had to take all of that information — the recognized entities, concepts, and keywords — and turn them into numbers so that Watson's deep-learning algorithms could pick them up. How would Watson determine if a player would be a boom or bust, or if he was injured, or if he might play meaningful minutes during a game that week? Answering those questions meant creating and refining models for deep learning, and conducting dozens of experiments to determine what might work. Finally, Baughman and his team had to build potential score spreads for every player each week, creating more than 1,000 simulations for each of the top 400 professional football players every day.
Running 1,000 simulations, 400 players, 1 team
The system debuted last November, letting fantasy players tap into Watson's insights via the ESPN website, where they could click on any individual football player to find a wealth of information and predictions on a tab labeled "Fantasy Insights with Watson." There, graphs suggested a range of projected scores for each player, with flags on each side labeling "boom" and "bust" percentages. Fantasy team managers could even run head-to-head player comparisons based on Watson's predictions.
In week 13, for instance, Watson compared two different quarterbacks, and found sentiment from news stories and social media postings favored one over the other. That Sunday, both quarterbacks got the win, but the quarterback with more positive sentiment blitzed his opponent — throwing five touchdown passes for 419 yards en route to victory.
For fantasy players, the system pulls in and pours over more unstructured data than the typical competitor could ever hope to analyze — who has time to read every article about every team and 400-plus players every day of the week? — determines the prevailing sentiment, and correlates that to probable outcomes. The hope is that by next season, Watson will also be able to help draft a fantasy team, and evaluate potential trades. It's a matter of continuing to hone and shape what Watson can learn — while also convincing fantasy owners to trust the computer as much as they trust themselves.
"Fantasy football players are vocal," Syken says with a laugh. "They're inherently skeptical of technology. You give them a tool and they still want to rely on their own human knowledge and intuition."
Moving beyond fantasy
That same thing can often be said for lawyers and doctors and businesspeople, but ideally, what both IBM developers and Watson are learning through fantasy football could be applied to all of those fields and more. If Watson can be taught football, it can also be taught to read, say, patient histories, helping to correlate patterns that medical professionals might not see as clearly otherwise. It could also sort through thousands of pages of legal or financial documents the way it worked through thousands of football-related articles and social-media posts-the tax-preparer H&R Block, for instance, is already using Watson to help it sort through changes in the tax code and make recommendations to its clients.
Really, any field that has huge volumes of information could potentially be aided by what Watson did for your fantasy team, according to Syken.
"These are the same kinds of trade-offs that businesspeople make every week," Syken says. "They're increasingly relying on the rows and columns and numbers of structured data, but they don't tap into, 'What are people saying about my business and my company?' If you really want to understand what's happening in your business, you have to be able to extract all that unstructured information as well."
Here's how Watson did it
Each week of the fantasy football season, Watson ingested and analyzed thousands of news stories, expert opinion pieces written by fantasy sports gurus, and hard numbers, including performance and injury reports for each of the players in the league. The system then correlated that information with traditional statistical data on over 1,900 players across 32 NFL teams.
The system calculated the potential upside and downside for each player, should a fantasy manager decide to start them on their team, assigning a "boom" or "bust" percentage — and the probability of a player performing with an injury. The result: fantasy managers could run quick calculations of their own using Watson: running risk vs. reward scenarios, comparing players head-to-head, and deciding on their best possible team week in and week out
Step One: Learning
A team of human annotators, data scientists, and developers using Watson Knowledge Studio, trained Watson to understand the concept of fantasy football. Watson learned the language of both NFL football and fantasy football, so the system could easily read and understand millions of news articles within the context of fantasy football — beginning the process of importing huge amounts of unstructured data.
Step Two: Comprehending
More than 90 gigabytes of unstructured text from historical fantasy football seasons were ingested into one model, while a second model was created based on information from football encyclopedias. The merged model was able to infer correct answers to football questions 97.96 percent of the time in analogy tests.
Step Three: Deeper learning
Using deep neural networks with more than 90 layers, the system was trained to determine if a player would boom, bust or play with an injury. For Watson, this wasn't a gut feeling, but the activation functions of the neural networks — a mixture of tanh and rectified linear units (ReLU). Deep player classifiers were developed, compared and scaled, with boom and bust percentages provided to fantasy players on the ESPN Fantasy Football analytics dashboard and on ESPN TV.
Step Four: Projecting a point spread
The last set of calculations asked Watson to assign point projections for each player. Using a support vector machine, Watson found the correct multiple regression function for each player position, factoring in each player's age, height, and other information to determine the point projection — and, after running simulations, a point spread.
The Fantasy Insights with Watson application was built from four Cloud Foundry applications running in three production IBM Cloud regions and one development IBM Cloud region. The python application was the machine learning engine pulling dozens of models from Object Storage — running in Dallas, Germany and London to ensure continuous availability. Watson is ingesting and analyzing thousands of news stories, expert opinion pieces by fantasy gurus, and reports on player injuries. The resulting insights are then correlated with traditional, statistical data on more than 1,900 players across all 32 NFL teams, to help fantasy owners decide which players to start every week.
The system calculates the potential upside and downside for each potential starter, analyzes whether a player will boom or bust, and determines the probability that a player will play with an injury. It lets a fantasy owner visualize the risk-and-reward scenarios, compare players, and field a more competitive team.